Research and Development of Line Loss Management and Load Forecasting System for Electric Power Enterprises Based on New Energy Consumption Technology Optimization

China has abundant electric power energy and belongs to a large electric power energy country [1]. In recent years, China has been committed to the development of electric power energy, to solve the drawbacks of the traditional petrochemical energy application process, to optimize the national energy structure, and to guarantee the security and stability of national energy supply [2–3]. After the production of electric energy, it needs to be transported to various regions by the electric power network to complete the transmission process. Power energy transmission process line loss is more serious, the need to take scientific management measures to improve line loss energy waste. It can be seen that the power enterprise line loss management is very necessary [4–6].

In the development process of electric power enterprises, due to technology, capital and many other reasons, resulting in insufficient construction of the power system, the network structure is not scientific, and new technologies can not be applied in a timely manner. Especially in the process of urban and rural network construction, the above problems are especially prominent, resulting in excessive power supply radius, extending the length of the conductor, and excessive resistance of the current in the line, leading to serious line losses [7–8]. In addition, the problems of insufficient equipment management, light and empty loads, and poor personnel quality have exacerbated the line loss problem [9–10].

Line loss is a problem that exists objectively in the energy management of electric power enterprises, but it is possible to reduce the impact of line loss through effective management measures, so line loss is an important symbol reflecting the management level of electric power enterprises. However, at present, China’s electric power enterprises line loss management awareness is insufficient, only line loss management stays on the surface of the understanding, and did not take effective measures to implement the line loss management in practice [11–12]. According to the causes of line loss, it can be seen that line loss occurs in all aspects of the power system, the need for comprehensive management, in order to improve the line loss problem. However, in the actual implementation process, due to the lack of awareness of line loss management, each department will focus on the main functions of the task, and do not realize the necessity of line loss management, resulting in line loss management has not been the attention of each department [13–15].

On the other hand, the stable supply and efficient utilization of electric power is the key to ensure the quality of life of the public. The wide application of smart grid technology prompts the power system to face huge data challenges and development opportunities. Power load forecasting and optimal control play an important role in ensuring the security of the power system [16–18].

Among the traditional methods of power load forecasting often most rely on historical data, meteorological information and holidays and other factors to assess the forecast, the staff can establish a mathematical model for power load forecasting. The traditional methods are difficult to ensure the accuracy of the electric load prediction results [19–21]. For this reason, big data technology has gradually been widely used in electric load prediction and optimization control. Big data technology has powerful data processing ability and excellent data model learning ability, which can improve technical support for the safe and stable operation of the power system [22–23].

In this paper, an active distribution network load forecasting model for electric power enterprises based on LSTNet is first constructed, and a subindustry load forecasting framework is designed by combining the federated learning method. Then, a novel low-carbon and economic dispatch method for electric-hydrogen hybrid energy storage of the power system is designed, and an arithmetic example analysis is performed to verify that the inclusion of new energy into system standby dispatch can optimize the comprehensive benefits of the system. Finally, on the basis of load forecasting and standby scheduling, this paper designs a line loss visualization and monitoring analysis system, which realizes the optimization of line loss management and load forecasting for power enterprises.

2

Problem of high proportion of new energy consumption being blocked

This chapter analyzes the problem of large-scale new energy consumption obstruction and its causes through the operating characteristics of wind power and photovoltaic, and lays the foundation for the research and development of line loss management optimization and load forecasting system for electric power enterprises.

2.1

Operational characteristics of wind power

Wind power is directly affected by wind resources, the output of stochastic fluctuation characteristics are obvious, in the night wind power output is larger, the wind power output is smaller during the day, and its trend is usually the opposite of the trend of load changes, with anti-peaking characteristics and stochasticity.

2.1.1

Randomness and volatility

In the case of a wind farm, for example, its average daily output is less than the rated output for most of the year, and rarely reaches the rated output, with the wind power output fluctuating randomly between zero and the installed capacity. The same wind farm has the problem that the total amount of wind resource is roughly the same in different days, but the proportion of output at different moments varies greatly, showing a strong randomness. At the same time, the wind power plant shows obvious “nightly” output characteristics, but its output not only has the stochasticity under the monthly time scale, but also has the stochasticity under the daily time scale, which reflects the obvious stochasticity and volatility.

2.1.2

Inverse FM characteristics

System load is usually in the trough period at night, the power grid by reducing the conventional power output to balance the source load side power. However, in the case of large-scale grid-connected wind power, the conventional power output adjustment speed is slower, and the unit is subject to the minimum technical output and other constraints, resulting in its downward adjustment of space is limited, coupled with the anti-peaking characteristics of the wind power will exacerbate the source-load power imbalance phenomenon at night. In order to ensure that the system can operate normally, it can only abandon the wind to maintain the source-load side power balance. Wind power as a reverse load and the system load superimposed to form the equivalent load, compared with the system load, the equivalent load curve peak-to-valley difference increases significantly.

2.2

PV Operating Characteristics

Unlike wind power output, PV power is strongly influenced by meteorological factors, i.e. irradiance, cloud motion and temperature and humidity, etc., and the same is true for PV power plant output, which is also characterized by randomness and volatility.

Photovoltaic power generation depends on the size of solar radiation, with “daytime and nighttime” characteristics, in the intensity of sunlight and temperature changes in weather conditions (cloudy, rain, snow, cloudy days, etc.), the PV power in a very short period of time to rapidly rise or fall, violent fluctuations occur. Due to the strong volatility and randomness of meteorological conditions, the output power of PV power plants is also strong randomness and strong volatility.

2.3

Operating characteristics of wind power

As the wind power anti-peaking effect makes the night load trough lower, photovoltaic in the evening peak power reduction makes the time load peak higher, as the scale of wind power access to the grid continues to expand, more and more caused by the system peak-valley difference does not reduce the problem of the anti-increase. In the case of abundant wind and light resources, wind and light power and system load superposition, resulting in system equivalent load peak and valley difference increases.

Wind farms in the midday moment of low power, and in the night with the early morning this stage of high power, photovoltaic power plant generation is the opposite of wind power, the two in different moments corresponding to the wind power big hair, photovoltaic small hair or wind power small hair, photovoltaic big hair just to make up for each other’s shortfall in the power, but due to the stochasticity of the solar energy and wind energy resources, fluctuation and inaccuracy of the predictability, there are photovoltaic and wind power at the same time, big hair or small hair at the same time, the moment, the wind power at the same time, the wind power at the same time. However, due to the stochastic, volatile and unpredictable nature of solar and wind resources, there are also moments when PV and wind power are simultaneously large or small, resulting in an increase in the peak-to-valley difference in equivalent load, so that large-scale grid-connection of wind power and PV will have a huge impact on the normal and stable operation of power systems.

2.4

Problem of high proportion of wind and solar power consumption blockage

2.4.1

Increased demand for power system peaking

From the operating characteristics of the high proportion of wind power, it can be seen that the large fluctuations and anti-peaking characteristics of wind power increase the peak demand on the system, and the “down-peaking” demand is particularly significant. In order to facilitate the quantitative analysis, can be illustrated by the system equivalent load. The specific calculation method is as follows:

Based on the wind power forecast processing and system load forecast value at each moment, the system equivalent load at each moment can be obtained $P_{e q u}^{t}$ : 1 $P_{e q u}^{t} = P_{e}^{t} - \sum_{i = 1}^{I} P_{w i n d, i}^{t} - \sum_{j = 1}^{J} P_{s o l a r, i}^{t}$

Where $P_{w i n d, i}^{t}$ denotes the predicted output of the ird wind turbine at moment t, $P_{s o l a r, i}^{t}$ . denotes the predicted output of the jth photovoltaic unit at moment t, and $P_{e}^{t}$ denotes the predicted value of system load at moment t.

Based on a regional power grid prediction data in the past few days as an example to analyze the peak demand of the grid, the peak demand of wind power access system is shown in Figure 1.

As can be seen from Figure 1, in the time period 11:00-14:00, wind and light resources form a complementary, but in the night time, the system load is reduced, when the wind power big hair makes the night system equivalent load trough is lower, PV in the evening peak power reduction makes the time period of the system equivalent load peak is higher. Therefore, wind power and photovoltaic power generation combined with the grid in some moments of the system equivalent load is lower, and the system equivalent load peak-valley difference compared to the system load peak-valley difference has increased significantly, increasing the power system on the peak shifting capacity requirements. In summary, with the increase in the proportion of wind and photovoltaic power sources connected to the grid, the system equivalent load peak to valley difference increases, making the power system’s peaking demand gradually increase.

2.4.2

Deficiencies in the “down-peaking” capacity of the grid

The current peak power mainly for coal power, hydropower and combined cycle gas power generation, due to hydropower units in the abundant water period does not participate in the peak, in the dry water period peak capacity by the water flow restrictions, combined cycle gas units by the constraints of the cogeneration model, the system peak capacity is smaller. When the proportion of new energy access to the grid increases, the proportion of adjustable power access decreases, making the power system “down-peaking” ability to reduce, resulting in the system generates wind and light abandonment. To analyze a regional power grid operation data on a certain day, new energy access ratio increased before and after the increase in the elimination of obstruction schematic shown in Figure 2 and Figure 3, respectively, by the system load minus the wind and light power output to get the equivalent load curve, in order to ensure that the system balance of power, lower than the lower limit of the system peaking of part of the wind and light power can not be eliminated by the system.

As can be seen from Figure 2, new energy access ratio increased before the existence of new energy consumption blocked power S1 and S2, in the S1 time, wind power big hair, and at this time the system load is in the trough, the system equivalent load is low, resulting in new energy consumption is blocked. In the S2 time, the light intensity increases significantly, the PV is large and reaches the peak output at 13:00, at this time the system equivalent load is also in the low valley, which makes the new energy appear to be blocked.

As can be seen from Figure 3, when the system load is unchanged, after the proportion of new energy access increases, the system equivalent load peak-valley difference is further increased, while the proportion of conventional power generation is reduced, the grid “down-peaking” resource shortage contradiction is significantly prominent, resulting in the new energy obstructed area S1, S2 increased to S3, S4, wind and solar power obstruction scale is further expanded. Expanded.

In summary, with the increase of new energy grid proportion, the equivalent load peak-to-valley difference further increases, the proportion of conventional power generation decreases, the system downshifting space is further reduced, the blocked new energy power generation increases, and the contradiction of insufficient grid “downshifting” resources becomes more and more prominent. Therefore, in order to meet the high proportion of new energy grid consumption of new energy goals, the need to make full use of grid peaking resources at the same time to find new peaking resources in order to form an effective consumption of new energy.

3

Active distribution network terminal load forecasting for electric utilities

3.1

LSTNet-based load forecasting model

Traditional recurrent neural networks (RNNs) suffer from short-term memory, which makes it difficult for RNNs to capture the information in the front of the sequence for longer sequences and face the problem of gradient vanishing during backpropagation [24]. Data-driven deep learning based models can learn the nonlinear patterns in load sequences and potential correlations between multivariate features (load, weather, time, and other features). LSTM introduces the structure of memory unit on the basis of RNN, and restricts the information flowing into the memory unit through the gating mechanism, so that the neural network not only remembers the previous information, but also selectively forgets the unimportant information, which solves the shortcomings of RNN [25]. The memory cell structure in LSTM consists of four parts: forgetting gate, input gate, memory cell and output gate. The forgetting gate is responsible for forgetting part of the information in the memory cell at the previous moment, the input gate is responsible for selectively inputting the input information into the memory cell structure, and the output gate is responsible for deciding what information is allowed to be output at the current moment.

However, limited by the network structure, LSTM cannot capture longer-term dependencies. Unlike LSTM, the structure of LSTNet (long- and short-term temporal time-series network) is shown in Fig. 4, which consists of three parts: convolutional layer, cyclic and cyclic-jumping layers, fully-connected layer, and autoregressive layer [26]. LSTNet is able to utilize the convolutional layer to extract multivariate short-term patterns and short-term dependencies among multivariate time-series from the input multivariate time series to extract the short-term patterns of the multivariables and the short-term dependencies among the multivariables, and then utilizes the recurrent and recurrent-jump layers to capture the long-term and longer-term dependencies among the multivariables.

In this paper, LSTNet is used as a load forecasting model, and the input of the model is a multivariate time series $X = {x^{(1)}, x^{(2)}, \dots, x^{(n)}}^{T} = {x_{1}, x_{2}, \dots, x_{T}} \in i_{i}^{n \times T}$ , where $x^{(k)} \in i_{i}^{T}$ , denotes the univariate time series of length T corresponding to the krd feature, and x_t ∈ iⁿ, denotes the multivariate time series containing n features at the moment of t. Here, there are a total of 27 features for the load, the weather and the time. In this paper, we predict x_T+1 based on (x₁, x₂, …, x_T}, and similarly, we predict x_T+h based on (x_1+h, x_2+h, …,x_T+h}. Memorize the observation y_T+h = x_T+h, and the predicted value Ŷ_T+h = f(x_1+h, x_2+h, …, x_T+h) of the model output. 1)

Convolutional Layer

The convolutional layer has d_c convolutional kernels of size (m,n), where m is the width of the convolutional kernel and n is the number of features. The kth convolution kernel scans the input sequence and outputs a h_k vector of length T: 2 $h_{k} = R E L U (W_{k} * X + b_{k})$ where * is the convolution operation and W_k and b_k are the weight coefficients and bias, respectively. The output matrix of the convolutional layer consists of d_c h_k -vectors of size (d_c, T), corresponding to weight coefficients and bias of W^c and b^c, respectively.

2)

Loop and loop-jump layers

The output of the convolutional layer is fed to both the loop layer and the loop-jump layer. The recurrent layer is a GRU model, and the unit structure of GRU consists of update gate and reset gate, which is simpler compared to LSTM, and each hidden layer reduces the number of matrix multiplications, which improves the training speed. The activation function is RELU. Formulas are shown in equations (3) to (6): 3 $r_{t} = σ (W_{r} [h_{t - 1}, x_{t}] + b_{r})$ 4 $z_{t} = σ (W_{z} [h_{t - 1}, x_{t}] + b_{z})$ 5 $c_{t} = R E L U (W_{c} [r_{t} * h_{t - 1}, x_{t}] + b_{c})$ 6 $h_{t} = (1 - z_{t}) * h_{t - 1} + z_{t} * c_{t}$

Where: r_t, z_t and c_t are the reset gate coefficients, update gate coefficients and candidate hidden layer coefficients at the moment of t, respectively; W_r, W_z and W_c are the weight coefficients of the reset gate, update gate and candidate hidden layer, respectively. b_r, b_z and b_c are the biases of the reset gate, update gate and candidate hidden layer, respectively. h_t–1 and h_t are the output values of the loop layer at moments t – 1 and t, respectively.

The loop-jump layer is also a GRU model, but p is introduced to capture the periodicity of the time series, as shown in Eqs. (7) to (10): 7 $r_{t} = σ (W_{r} [h_{t - p}, x_{t}] + b_{r})$ 8 $z_{t} = σ (W_{z} [h_{t - p}, x_{t}] + b_{z})$ 9 $c_{t} = RELU (W_{c} [r_{t} * h_{t - p}, x_{t}] + b_{c})$ 10 $h_{t} = (1 - z_{t}) * h_{t - p} + z_{t} * c_{t}$ where p is the number of skipped hidden units and h_t–p is the output value of the loop-jump layer at moment t – p. In order to distinguish between the loop layer and the loop-jump layer, let $h_{t}^{R}$ be the output value of the loop layer at moment t and $h_{t}^{s}$ be the output value of the loop-jump layer at moment t.

3)

Fully connected layer and autoregressive layer

The fully connected layer receives the outputs of the loop layer and the loop-jump layer as shown in the following equation: 11 $h_{t}^{D} = W^{R} h_{t}^{R} + \sum_{i = 0}^{p - 1} W_{i}^{S} h_{t - i}^{S} + b^{D}$ where $h_{t}^{D}$ is the output value of the fully connected layer at moment t, $h_{t - i}^{S}$ is the output value of the loop-jump layer at moment t – i, W^R and $W_{i}^{s}$ are the weight coefficients of the loop layer at moment t and loop-jump layer at moment i, respectively, and b^D is the bias of the fully connected layer.

The autoregressive layer is responsible for providing the linear component, which helps to improve the prediction efficiency in large-scale data. It is shown in the following equation: 12 $h_{t, i}^{L} = \sum_{k = 0}^{q^{a r} - 1} W_{k}^{a r} y_{t - k, i} + b^{a r}$ where $h_{t}^{L}$ is the output of the autoregressive layer, q^ar is the size of the input window acting on the input sequence, and W^ar and b^ar are the weight coefficients and bias of the autoregressive layer, respectively.

The final output prediction is the superposition of the output of the neural network and the linear component of the autoregressive layer: 13 $y_{t} = h_{t}^{D} + h_{t}^{L}$ where y_t is the prediction result at moment t. Remember that the weight coefficient corresponding to the convolutional layer is W^C and the bias is b^c, the weight coefficient corresponding to the cyclic layer is W^R and the bias is b^R, the weight coefficient corresponding to the cyclic-jump layer is W^S and the bias is b^S, the bias of the fully-connected layer is b^D, and the weight coefficient corresponding to the autoregressive layer is W^ar and the bias is b^ar. Collectively, we will write down weight coefficients W = {W^C, W^R, W^S, W^ar} and bias b = {b^C, b^R, b^S, b^D, b^ar} for the layers in the LSTNet as: 14 $ω \in {W, b}$ where ω represents the weight coefficients and bias of the neurons in each layer, i.e., the model parameters of LSTNet.

3.2

Federated Learning-based Load Forecasting by Subsector

In order to protect the user reading meter data while accessing the load forecasting model, this paper uses federated learning to design a subsector load forecasting framework. Federated learning is a machine learning based on distributed datasets, which separates the process of directly accessing training data from the process of training models. Federated learning consists of two processes: model training and model inference. In the model training process, multiple clients train locally with their own datasets, and the server collects the model parameters to do the global model update, and each client continues to train based on this update, and finally obtains a model that can be shared by multiple parties. Model inference is the application of the trained federated learning model to new data. In view of the overlap of data from different clients in both feature space and sample space, federation learning is categorized into horizontal federation learning, vertical federation learning and federation migration learning. Among them, horizontal federation learning applies to the situation where there is overlap in feature space between clients and the sample space is different. The applicable conditions of horizontal federation learning are shown in equation (15): 15 $X_{i} = X_{j}, I_{i} \neq I_{j}, \forall D_{i}, D_{j}, i \neq j$

Where D_i and D_j. are the datasets of Client i and Client j respectively, X_i and X_j are the feature spaces of Client i and Client j respectively, and I_i, and I_j are the sample spaces of Client i and Client j respectively.

In the application scenario of this paper, different power end-users in the same industry each have their own datasets with different samples, but the samples correspond to similar features (load shape, weather information, time information). The active distribution grid operator can obtain a global load forecasting model for the industry through horizontal federated learning and provide the model to the electricity selling company, which is able to forecast the load of the agent’s customers without collecting access to its agent’s end-user load meter reading data for training.

3.2.1

Federal Learning Open Source Library

FedML is a federated learning library developed by USC, which supports local deployment, distributed deployment and mobile deployment, and it has built-in federation optimization algorithms such as FedAvg, FedOpt, FedNas and supports customization.

The federated learning training in this chapter is implemented based on FedML library, which consists of FedML-API and FedML-core, and FedML-core consists of distributed communication module and model training module, in which the distributed communication module is based on MPI (message passing interface) to realize distributed computing, and the model training module is based on PyTX (message passing interface) to realize distributed computing. The distributed communication module is based on the MPI (message passing interface) network transmission interface to realize distributed computing, and the model training module is based on the PyTorch deep learning library.FedML-API is the high-level API of FedML-core, which adopts the worker-oriented design pattern, and it can be programmed with the behaviors of workers and customize the messages exchanged by each worker, thus improving the flexibility of distributed computing.

3.2.2

Federated training process based on FedAvg algorithm

Federated learning is based on a federated optimization algorithm to obtain a global model.The FedAvg algorithm is the first horizontal federated optimization algorithm that runs the gradient descent algorithm iteratively by running it in parallel on multiple user-side clients and sending it back to the clients to continue training after aggregating the parameters collected by a central server in each round of interactions [27].

The federation training information flow based on the FedAvg algorithm is shown in Fig. 5. The users participating in federated learning are response users, and the model parameters are the weight coefficients or bias of the neurons in each layer of LSTNet. The active distribution network operator sets up a central server for aggregating the local model parameters uploaded by the response users, and the response users use the local dataset to conduct local training based on the LSTNet load prediction model without sharing the dataset with each other. After the local training reaches the set number of local training batches, the responding users upload the local model parameters to the central server. After collecting the local model parameters uploaded by all the responding users participating in the federated training, the central server generates global model parameters based on the FedAvg algorithm and sends them back to each responding user. Responding users update their local model parameters with the global model parameters and continue training, repeating the process until the set number of interaction rounds is reached.

Let interaction round t = 0,2,…, T be the number of times the response users transmit local model parameters to the central server located at the active distribution network operator, local training batch e = 1,2,…, E be the number of local training iterations of the LSTNet load forecasting model based on LSTNet for the response users, and model parameter ω be shown in Equation (14), and ω_t be the weight coefficients or bias of the neurons of each layer of the LSTNet after the local training batch reaches E in the tth round of interaction. First, the central server located at the active distribution network operator sends the initialized model parameters ω₀ to all responding users. After receiving the model parameters, the response users train the LSTNet model locally based on the gradient descent algorithm. In the tth round of interaction, the response users k update: 16 $ω_{t}^{k} = ω_{t - 1}^{k} - η \nabla f (ω_{t - 1}^{k})$ and then send it to the central server located at the active distribution network operator. The central server, after collecting all the model parameters $ω_{t}^{k} (k = 1, 2, \dots, K)$ updated by the responding users, does aggregation of the parameters to generate global parameters: 17 $ω_{t} = \sum_{k = 1}^{K} \frac{n_{k}}{n} ω_{t}^{k}$ where n_k is the number of samples from the responding user k and n is the total number of samples from all responding users. The central server sends global parameters ω_t to all responding users, and the responding users update local parameters based on the global parameters: 18 $ω_{t}^{k} = ω_{t}$

The training then continues until the T st interaction is completed.

3.2.3

Federated Learning-based Load Forecasting Framework for Subsectors

The overall architecture of the federated learning-based sub-sector load is shown in Fig. 6, and the main body includes the active distribution grid operator, the central server set up by the active distribution grid operator, the users, and the power sales company, and the whole process is divided into the following seven steps: 1)

The active distribution grid operator sends a training request to the user corresponding to a target industry.

2)

Considering that the user needs to complete the training locally, the user can decide whether or not to participate in the federal training. The participating users are the responding users, which preprocess their dataset locally and input the dataset into the LSTNet load forecasting model for local training, and pass the neuron parameters of each layer of LSTNet to the central server after completing one round of training.

3)

The central server aggregates the local model parameters passed by the responding users based on the FedAvg algorithm to generate global model parameters and passes the global model parameters to the responding users, and the central server obtains the industry global model after several interactions.

4)

The central server passes the industry global model to the active distribution network operator.

5)

The active distribution network operator distributes rewards according to the degree of contribution of the responding users to the industry global model.

6)

The electricity selling company submits a model demand request to the active distribution grid operator based on the desired industry.

7)

The Active Distribution Grid Operator returns the required industry global model and receives the revenue.

3.3

Analysis of the experimental effect of load forecasting models

In order to verify the reliability of the LSTNet grid load forecasting model based on federated learning, this paper analyzes its performance. The dataset in this paper is collected from the open source data websites Open Power System Data and Open Energy Data Initiative, which contains renewable energy output information, tariff information, and customer load information, and only the customer loads are selected for load forecasting in this chapter.

3.3.1

Impact of different learning rates on global models

In this paper, we first test the effect of different learning rates of local clients on the global model. The loss of the global model under different learning rates is shown in Fig. 7.

Some observations can be drawn from Fig. 7. First, with the increasing number of global aggregation rounds, the losses of the global model under different learning rates are decreasing. Second, in the late stage of global model training, the dark cyan line and the orange line are very close to each other in terms of global model loss values, but there is still a certain gap between the two as can be seen from the local zoomed-in plots. Finally, the dark purple line has the lowest global model loss compared to the other two lines. Therefore, according to the specific effect demonstrated in Fig. 7, the learning rate corresponding to the dark purple line, i.e., α = 1 × 10⁻⁴, is the learning rate that performs better in global model load prediction.

In order to further demonstrate the excellent performance of the global load forecasting model at the learning rate α = 1 × 10⁻⁴ compared to the other learning rates, this paper randomly selects one electricity user in each of the three types of electricity user groups and tests the prediction effect of the global load forecasting model on its local dataset. The results of the customer load prediction are shown in Fig. 8, where (a) to (c) denote industrial, commercial, and residential customers, respectively. The left side of the red dashed line in the figure shows Decoder’s prompt information, and the right side of the red dashed line shows the actual load prediction results.

According to Fig. 8, the observation results can be derived as follows: 1)

For industrial users, the global load forecasting model shows some deviations for the initial and ending stages of the forecast at all three learning rates, but compared to the other learning rates, the deep purple line, i.e., α = 1 × 10⁻⁴, predicts the trend of load fluctuations better. Next, in the mid-phase of load forecasting, the global load forecasting model predicts the trend of load fluctuation better at all three different learning rates, but the deep purple line predicts the specific load values better compared to the other learning rates.

2)

For commercial customers, the model has some deviation in the initial stage of forecasting under all three learning rates, but performs well in the end stage of forecasting. And in the middle stage of load forecasting, the model is able to predict the trend of load fluctuation better under all three different learning rates, but the deep purple line is better than the other learning rates.

3)

For residential customers, the prediction effect of the global load forecasting model under the three different learning rates is similar to that of commercial customers, i.e., there is a certain degree of deviation in the early stage of forecasting, but the prediction effect in the later stage is relatively better. However, in the middle stage of load forecasting, the global load forecasting model produces a certain degree of deviation in the prediction of load under the three different learning rates. For the dark cyan line, i.e., α = 1 × 10⁻³, the prediction of the load fluctuation trend and specific load values is completely lost in the middle stage of prediction. For the orange line, α = 1 × 10⁻⁵, the prediction of load fluctuation trend in the middle of the forecast is better than that of the dark cyan line, but the prediction of the specific load value still has some problems. For the dark purple line, although there is a certain degree of deviation in the middle of the forecast, its prediction of the load fluctuation trend and the specific load value is better than that of the dark cyan line and the orange line.

3.3.2

Comparison of the predictive effectiveness of the global model and the local model

In order to further test the prediction effect of the global load forecasting model, this paper trained three different LSTNet load forecasting models individually on the local datasets of the three types of power user groups. Subsequently, this paper selects the best-performing global load forecasting model to compare with the above three local models in terms of the specific effects of forecasting, and the results of the forecasting comparison for industrial, commercial, and residential users are shown in Fig. 9’s (a)~(c), respectively.

For industrial users, the performance of the global load forecasting model and the local load forecasting model in the middle of the forecast is similar, but in the beginning and the end of the forecast, the global model fits the trend of load fluctuation better than the local model, and there is still a certain gap between the two models in terms of the forecasting effect of specific load values.

For commercial users, the performance of the global model and the local model in the mid- and late-periods of the forecast is similar, but in the initial stage of the forecast, the local model outperforms the global model in terms of the trend of load fluctuations and the prediction of specific load values. The global model has a certain prediction effect on the load fluctuation trend, but there is still a certain gap in the prediction of specific load values.

For residential customers, there is not much difference between the global model and the local model in the late stage of forecasting, but there is still a gap between the two models in the initial stage and the middle stage of forecasting. First, the global model outperforms the local model in the initial stage of prediction. Secondly, from Fig. 9(c), it can be seen that in the middle stage of forecasting, the local model performs poorly in predicting the load fluctuation trend. On the contrary, the global model is stronger than the local model in predicting the load fluctuation trend, although there is a certain gap in the prediction of specific values.

3.3.3

Comparative analysis of load forecasting metrics of models

Finally, this paper analyzes the global load forecasting model using three common regression task evaluation metrics. Among them, MAE is the mean absolute error, which is calculated by the formula: 19 $M A E = \frac{1}{N} \sum_{i = 1}^{N} | y_{i}^{l a b e l} - y_{i}^{p r e d} |$

It measures the magnitude of the mean error and does not take into account positive or negative.

MSE is the Minimum Mean Square Error, which is a common loss function and is calculated as: 20 $M S E = \frac{1}{N} \sum_{i = 1}^{N} {(y_{i}^{l a b e l} - y_{i}^{p r e d})}^{2}$

It measures the sum of the squares of the errors of the predicted values and the true labels.

RMSE is the Root Mean Square Error, which is calculated as: 21 $R M S E = \sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(y_{i}^{l a b e l} - y_{i}^{p r e d})}^{2}}$

The lower values of the above three evaluation indicators indicate the stronger fitting ability of the regression model.

The indicator values of the global load forecasting model for the load forecasting tasks of the three types of power consumers, and the indicator values of the LSTNet model trained separately on the local datasets of each type of power consumers are shown in Table 1.

Table 1.

Comparison of model load prediction index

Model	The global model			The local model
Index	MAE	MSE	RMSE	MAE	MSE	RMSE
Industrial Dataset	0.108	0.043	0.163	0.069	0.022	0.136
Commercial Dataset	0.239	0.084	0.242	0.176	0.041	0.197
Residential Dataset	0.358	0.233	0.469	0.287	0.163	0.407

According to Table 1, it can be seen that all the metrics of the LSTNet model trained separately on the local dataset take lower values than the global load forecasting model, so its fitting ability is stronger than that of the global load forecasting model. However, the individually trained LSTNet model cannot realize the load prediction of multiple classes of users, while the global load forecasting model can realize the task of predicting the load of multiple classes of electric power users due to the parameter averaging in the federated learning process that makes it learn the laws of load data distribution of different classes of users. In addition, although the global load forecasting model is larger than the individually trained local model in all the indexes, the difference in the values of each index is not large, and the performance of the global load forecasting model is very close to that of the local load forecasting model in terms of the actual forecasting effect of the global model.

4

Low-carbon economic dispatch of new power systems considering hybrid electric-hydrogen energy storage

In order to increase the amount of new energy connected to the grid and reduce the difficulty of new energy consumption caused by the backup provided by thermal power units, this chapter proposes a hybrid electric-hydrogen energy storage to participate in a new type of power system low-carbon and economic scheduling method.

4.1

First-stage dispatch model taking into account hybrid electric-hydrogen energy storage

4.1.1

Objective function

The first stage minimizes the load variance as an objective function to minimize the peak-to-valley difference in the load curve after demand response, and the objective function is shown below: 22 $f_{1} = \min {\sum_{i = 1}^{T} {[P_{1, t}^{P D R} - (\sum_{i = 1}^{T} P_{1, t}^{p r e}) / T]}^{2}}$

Where $P_{1, t}^{P D R}$ represents the load magnitude in the t period after the demand response and $P_{1, t}^{p r e}$ represents the load magnitude in the t period before the demand response.

4.1.2

Constraints

Demand response transfers load by changing the way customers use electricity, with no change in the total amount of load before or after the transfer: 23 $\sum_{i = 1}^{T} P_{1, t}^{p r e} = \sum_{i = 1}^{T} P_{1, t}^{P D R}$

Demand response cannot be adjusted indefinitely, and the adjustment process will change the way the customer uses electricity. Therefore, the demand response curve should be formed by considering the constraints of satisfaction with the way of using electricity and satisfaction with the expenditure on electricity, which are shown below: 24 ${\begin{matrix} 1 - \frac{\sum_{t = 1}^{24} | λ_{p, t} |}{\sum_{t = 1}^{24} P_{1, t}^{p r e}} \geq M_{\min}^{μ} \\ 1 + \frac{\sum_{t = 1}^{24} (D_{t}^{p D R} P_{1, t}^{P D R} - D_{t} P_{1, t}^{p r e})}{\sum_{t = 1}^{24} P_{1, t}^{p r e} D_{t}} \geq M_{\min}^{p} \end{matrix}$

Where, λ_p,t is the amount of load change in each time period, D_t is the price of electricity in the t time periods before PDR, $D_{t}^{P D R}$ is the price of electricity in the t time periods after PDR optimization, $M_{\min}^{μ}$ is the minimum value of customer satisfaction with the way of using electricity, and $M_{\min}^{p}$ is the minimum value of satisfaction with the expenditure on electricity.

4.2

Second Stage Dispatch Model Taking into Account Hybrid Electric-Hydrogen Energy Storage

4.2.1

Objective function

The second stage is the day-ahead scheduling stage, which is aimed at optimizing the comprehensive system cost, including thermal power unit operation cost, storage construction operation and maintenance cost, carbon trading cost, standby cost, and energy loss cost. Namely: 25 $f_{2} = \min (C_{o p} + C_{e s} + C_{y} + C_{b} + C_{q})$

Where, f₂ is the comprehensive system cost, C_op is the thermal power unit operation and maintenance cost, C_es is the average daily cost of energy storage, C_y is the carbon trading cost, C_b is the standby cost, and C_q is the energy loss cost. 1)

Thermal power unit operation and maintenance costs

The first thermal power unit is a carbon capture unit, and the operation and maintenance cost of the unit includes the power generation cost of the thermal power unit, the start-stop cost and the solution loss cost of the carbon capture unit: 26 $C_{o p} = C_{f d} + C_{q t} + C_{r y}$

Where C_fd represents the cost of thermal unit power generation, C_qt represents the cost of thermal unit startup and shutdown, and C_ry represents the cost of carbon capture unit solution loss.

Thermal power unit power generation cost: 27 $C_{f d} = \sum_{t = 1}^{T} (\sum_{i = 1}^{n} (a_{i} P_{H i, t}^{2} + b_{i} P_{H i, t} + c_{i})$

Where T indicates the scheduling cycle. n indicates the number of thermal power units. P_Hi,d denotes the power generated by the thermal power unit during the t th period. a_i, b_i and c_i represent the cost coefficients of the ith thermal power unit.

Thermal power unit startup and shutdown cost: 28 $C_{q t} = \sum_{t = 1}^{T} \sum_{i = 1}^{n} [U_{i, t} (1 - U_{i, t - 1}) + U_{i, t - 1} (1 - U_{i, t})] S_{g, i}$

Where, U_i,t represents the i thermal unit t start-stop status and S_g,i represents the unit i start-stop cost.

Solution loss cost: 29 $C_{r y} = \sum_{t = 1}^{T} K_{s} φ E_{J c, t}$ where K_s represents the ethanolamine solution cost factor, φ represents the solution operating loss factor, and E_Jc,t represents the CO₂ amount of carbon captured by the carbon capture unit t during the time period.

2)

Average daily cost of energy storage

The energy storage in this chapter is electric-hydrogen hybrid energy storage, and the average daily cost of energy storage is divided into two parts: construction cost as well as operation cost. Namely: 30 ${\begin{array}{l} C_{e s} = C_{i n v} + C_{s o p} \\ C_{i n v} = (\sum_{j \in M_{t}} (\frac{γ {(1 + γ)}^{Y_{j}}}{{(1 + γ)}^{Y_{j}} - 1} c_{j}^{i n v} P_{j}^{N}) + \sum_{i \in M_{t}} \frac{γ {(1 + γ)}^{Y_{i}}}{{(1 + γ)}^{Y_{t}} - 1} c_{t}^{i n v} E_{t}^{N}) / 365 \\ C_{s o p} = \sum_{t = 1}^{T} (c_{e l} P_{e l, t} + c_{f c} P_{f c, t} + c_{d} (P_{c h a, t} + P_{d i s, t})) \end{array}$ where C_inv represents the daily construction cost of energy storage. C_sop denotes the hydrogen energy storage daily operation and maintenance cost. M_t represents the set of energy storage power devices, including electrolyzer, fuel cell, and storage battery. N_t represents the energy storage capacity equipment set, including hydrogen storage tank capacity and electric storage battery capacity. γ represents the discount rate, with a value of 0.05. Y_i and Y_j represent the useful life of equipment in categories i and j. $c_{i}^{i n v}$ , $c_{j}^{i n v}$ indicates the unit investment cost coefficients of the equipment in categories i and j. $P_{j}^{N}$ indicates the power of equipment in category j. $E_{i}^{N}$ indicates the capacity of equipment in category i. c_d indicates the battery unit charging and discharging cost coefficient. P_cha,t, P_dis,t indicates t hours of battery charging and discharging power.

3)

Carbon trading costs: 31 $C_{y} = \sum_{t = 1}^{T} c_{t j} (E_{c}^{a l l} - \sum_{i = 1}^{n} σ P_{H i, t})$

Where c_ij represents the carbon trading price, $E_{c}^{a l l}$ represents the total carbon emissions of the system, and σ represents the carbon trading quota.

4)

Standby cost 32 $C_{b} = \sum_{k = R_{m}} \sum_{t = 1}^{T} (c_{k}^{u p} P_{k, t}^{u p} + c_{k}^{d o w n} P_{k, t}^{d o w n})$ where R_m is the set of equipment providing spinning reserve, including conventional thermal units, carbon capture units, and hydrogen energy storage systems. $c_{k}^{u p}$ and $c_{k}^{d o w n}$ are the cost coefficients for upper and lower rotating standby for category k equipment, respectively. $P_{k, t}^{u p}$ and $P_{k, t}^{d o w n}$ are the upper and lower rotating standby capacities provided by the equipment of category k in time period t, respectively.

5)

Power loss cost: 33 $C_{q} = \sum_{t = 1}^{T} c_{q} (P_{w, t}^{q} + P_{v, t}^{q} + (1 - η_{e l}) P_{e l, t} + (1 - η_{f c}) \frac{P_{f c, t}}{η_{f c}})$

Where $P_{v, t}^{q}$ is the abandoned wind power in time period t, $P_{v, t}^{q}$ is the abandoned light power in time period t, and c_q is the penalty cost coefficient for electricity abandonment.

4.2.2

Constraints

1)

Power balance constraints: 34 $\sum_{i = 1}^{n} P_{J i, t} + P_{w, t} + P_{v, t} + P_{f c, t} + P_{d i s, t} = P_{1, t}^{P D R}$

Where P_Ji,t represents the net output power of the i nd thermal power unit, P_w,t - the grid-connected power of wind power in the tth period, and the grid-connected power of photovoltaic power in the P_v,tth - tth period.

2)

Thermal power unit start-stop constraint: 35 ${\begin{array}{l} \sum_{k = t}^{t + T_{s} - 1} (1 - U_{i, k}) \geq T_{s} (U_{i, t - 1} - U_{i, t}) \\ \sum_{k = t}^{t + T_{o} - 1} U_{i, k} \geq T_{o} (U_{i, t} - U_{i, t - 1}) \end{array}$ where T_s and T_o are the minimum shutdown and start-up times, respectively.

3)

Climbing constraints for thermal power units: 36 ${\begin{array}{l} P_{H i, t} - P_{H i, t - 1} + U_{i, t - 1} (P_{H i}^{\max} - R_{i}^{u p}) + U_{i, t} (P_{H i}^{\max} - P_{H i}^{\min}) \leq P_{H i}^{\max} \\ P_{H i, t - 1} - P_{H i, t} + U_{i, t} (P_{H i}^{\min} - R_{i}^{d o w n}) + U_{i, t - 1} (P_{H i}^{\max} - P_{H i}^{\min}) \leq P_{H i}^{\max} \end{array}$

Where $R_{i}^{u p}$ and $R_{i}^{d o w n}$ represent the upward and downward climbing rates of the i rd thermal power unit, respectively.

4)

Carbon capture unit constraints

Carbon capture unit constraints include carbon capture unit energy flow constraints, carbon capture unit carbon dioxide flow constraints, and carbon capture unit liquid storage tank constraints.

(1)

Energy flow constraints: 37 ${\begin{array}{r} P_{H c, t} = P_{J c, t} + P_{B, t} \\ P_{B, t} = P_{r, t} - P_{D} \\ P_{r, t} = λ E_{J c, t} \end{array}$

Where P_{Hc, t} is the power generation output of the CCS unit in time period t, P_{Jc, t} is the net output of the CCS unit in time period t, P_B,t is the energy consumption of the CCS unit, P_D is the fixed energy consumption of the CCS unit, P_r,t is the energy consumption of the CCS unit in operation in time period t, λ is the energy consumed per unit of CO₂ captured, and E_Jc,t is the amount of CO₂ captured in time period t by the CCS unit.

(2)

CO₂ flow constraint: 38 ${\begin{array}{r} E_{c, t} = m_{c} P_{H c, t} \\ E_{J c, t} = η δ_{t} E_{c, t} + E_{v, t} \\ 0 \leq E_{J c, t} \leq γ_{c} η m_{c} P_{H c, t} \\ V_{c a, t} = E_{v, t} M_{v} / (M_{C O_{2}} θ ρ_{v} δ_{v}) \end{array}$

Where E_c,t is the amount of CO₂ produced by the carbon capture unit at time t, m_c is the carbon emission intensity of the carbon capture unit, η is the carbon capture efficiency, δ_t is the flue gas split ratio at time t, γ_c is the maximum operating condition factor, E_v,t is the amount of CO₂ supplied by the tank at time t, V_ca,t is the volume of solution required to supply CO₂ from the tank at time t, M_v is the molar mass of ethanol amine solution, M_CO2 is the molar mass of carbon dioxide, θ is the regeneration tower resolution, ρ_v is the concentration of the alcohol solution, and δ_v is the density of the amine solution.

(3)

Reservoir constraints: 39 ${\begin{array}{r} V_{F, t} = V_{F, t - 1} - V_{c a, t} \\ 0 \leq V_{F, t} \leq V_{c a}^{N} \\ V_{F, 0} = V_{F, 24} \\ V_{P, t} = V_{P, t - 1} - V_{c a, t} \\ 0 \leq V_{P, t} \leq V_{c a}^{N} \\ V_{P, 0} = V_{P, 24} \end{array}$

Where V_F,t and V_P,t are the amount of CO₂ produced by the carbon capture unit in time period t, m_c is the carbon emission intensity of the carbon capture unit, η is the carbon capture efficiency, δ_t is the flue gas split ratio in time period t, γ_c is the maximum operating condition factor, E_v,t is the amount of CO₂ provided by the reservoir in time period t, V_ca,t is the volume of solution required to provide carbon dioxide in the reservoir in time period t, M_v is the molar mass of ethanolaminesolution, M_CO2 is the molar mass of carbon dioxide, θ is the regeneration tower resolution, ρ_v is the alcohol solution concentration, and δ_v is the amine solution density.

Where V_F,t and V_P,t are the volume of solution in the liquid-rich and liquid-poor reservoirs at time t, respectively. $V_{c a}^{N}$ is the capacity of the reservoir. V_F,o, V_P,o are the initial time rich liquid storage tank and poor liquid storage tank solution volume. V_F,24 and V_P,24 are the solution volumes in the liquid-rich and liquid-poor tanks at the end of the period, respectively.

5)

New energy output constraints

Wind power output constraints: 40 ${\begin{array}{l} P_{w, t}^{p r e} = P_{w, t} + P_{e l, t}^{w} + P_{c h a, t}^{w} + P_{w, t}^{q} \\ 0 \leq P_{w, t} \leq P_{w, t}^{p r e} \end{array}$ where $P_{e l, t}^{w}$ denotes wind power hydrogen production power and $P_{c h a, t}^{w}$ denotes wind power charging power.

Photovoltaic power output constraint: 41 ${\begin{array}{l} P_{v, t}^{p r e} = P_{v, t} + P_{e l, t}^{v} + P_{c h a, t}^{v} + P_{w, t}^{q} \\ 0 \leq P_{v, t} \leq P_{v, t}^{p r e} \end{array}$ where $P_{e l, t}^{v}$ represents the PV hydrogen production power and $P_{c h a, t}^{v}$ represents the PV charging power.

6)

Energy storage constraints

The energy storage constraints include the constraints related to electrochemical energy storage and the constraints related to hydrogen energy storage.

In order to maintain the operating life of the electrochemical energy storage, the maximum charging and discharging power of the electrochemical energy storage has upper and lower limits, and the phenomenon of over-charging or over-discharging is not allowed. In addition, the electrochemical energy storage state is described using the battery charge state: 42 ${\begin{array}{l} 0 \leq P_{c h a, t} \leq P_{d}^{\max} \cdot U_{t}^{b e s s} \\ 0 \leq P_{d i s, t} \leq P_{d}^{\max} \cdot (1 - U_{t}^{b e s s}) \\ S_{t} = S_{t - 1} + \frac{P_{c h a, t}}{E_{d}^{N}} - \frac{P_{d i s, t}}{E_{d}^{N}} \\ S_{0} = S_{24} \\ S_{\min} \leq S_{t} \leq S_{\max} \end{array}$

Where: $P_{d}^{\max}$ indicates the maximum charging power of the battery. $U_{t}^{b e s s}$ indicates the battery state, a value of 1 indicates charging and a value of 0 indicates discharging. $E_{d}^{N}$ indicates the installed capacity of electrochemical energy storage. S_t indicates the battery charge state. S_min indicates the minimum charge state. S_max 6 indicates the maximum charge state. 43 ${\begin{array}{l} E_{h s, 0} = E_{h s, T} \\ 0.1 P_{e l, t}^{N} \leq P_{e l, t} \leq P_{e l, t}^{\max} \\ 0 \leq P_{f c, t} \leq P_{f c, t}^{\max} \\ E_{h s}^{\min} \leq E_{h s, t} \leq E_{h s}^{\max} \end{array}$

7)

Rotating standby constraints

(1)

Thermal unit standby constraints: 44 $0 \leq P_{i, t}^{u p} \leq \min (U_{i, t} P_{H i}^{\max} - P_{J, t}, Δ t R_{i}^{u p})$ 45 $0 \leq P_{i, t}^{d o w n} \leq \min (P_{J i, t} - U_{i, t} P_{H i}^{\min}, Δ t R_{i}^{d o w n})$

Where $P_{i, t}^{u p}$ and $P_{i, t}^{d o w n}$ represent the upper and lower rotating standby capacities assumed by the ith thermal power unit, respectively.

(2)

Hydrogen storage standby constraint: hydrogen storage has equivalent rotating standby characteristics, which consists of two parts: the equivalent rotating standby characteristics of the electrolyzer in the power generation stage and the equivalent rotating standby characteristics of the fuel cell in the power consumption stage.

Electrolyzer equivalent rotating standby constraint: 46 ${\begin{array}{l} 0 \leq P_{e l, t}^{u p} \leq X_{t} (P_{e l, t}, 0.1 P_{e l, t}^{N}) \\ 0 \leq P_{e l, t}^{d o w n} \leq \min [X_{t} (P_{e l}^{N} - P_{e l, t}), X_{t} \frac{E_{h s}^{\max} - E_{h s, t}}{η_{e l}}] \end{array}$

Where: $P_{e l, t}^{u p}$ and $P_{e l, t}^{d o w n}$ denote the upper spinning reserve capacity and lower spinning reserve capacity provided by the electrolyzer in time period t, respectively. X_t denotes the start-stop state of the electrolyzer equipment, a value of 1 means the equipment is started, and a value of 0 means the equipment is shut down.

Fuel cell equivalent rotating standby constraint: 47 ${\begin{array}{l} Y_{t} = 1 \\ 0 \leq P_{f c, t}^{u p} \leq \min [P_{f c}^{N} - P_{f c, t}, (E_{h s, t} - E_{h s}^{\min}) η_{f c}] \\ 0 \leq P_{f c, t}^{d o w n} \leq Y_{t} P_{f c, t} \end{array}$

Where: $P_{f c, t}^{u p}$ and $P_{f c, t}^{d o w n}$ denote the up-rotation standby capacity and down-rotation standby capacity provided by the fuel cell at time t, respectively. Y_t denotes the start-stop state of the fuel cell equipment, which is 1 for equipment startup and 0 for equipment shutdown.

In the shutdown state there is still 50% of the capacity with fast response capability. The fuel cell equivalent rotating reserve constraint in the shutdown state is: 48 ${\begin{array}{l} Y_{t} = 0 \\ 0 \leq P_{f c, t}^{u p} \leq \min [0.5 P_{f c}^{N}, (E_{h s, t} - E_{h s}^{\min}) η_{f c}] \\ P_{f c, t}^{d o w n} = 0 \end{array}$

(3)

System standby constraints: 49 $C_{r} {\sum_{i = 1}^{n} P_{J i, t} + P_{w, t} + P_{v, t} + Δ P_{w, t}^{ε} + Δ P_{v, t}^{ε} + P_{f c, t} + + P_{d i s, t} + \sum_{k = N_{m}} P_{k, t}^{u p} \geq P_{l, t}^{p D R} + Δ P_{1, t}^{ε}} \geq α$

Converting rotational standby constraints under triangular fuzzy parameters to clear equivalence classes: 50 $C_{r} {\sum_{i = 1}^{n} P_{J i, t} + P_{w, t} + P_{v, t} + Δ P_{w, t}^{ε} + Δ P_{v, t}^{ε} + P_{f c, t} + P_{d i s, t} - \sum_{k = N_{m}} P_{k, t}^{d o w n} \leq P_{1, t}^{p D R} + Δ P_{1, t}^{ε}} \geq α$ where C_r{ } is the confidence expression and α is the confidence that the system rotational standby constraint is satisfied.

8)

Trend constraints

The power network uses DC tidal current constraints, this part has been studied in more literature and will not be repeated here.

4.3

Model solving framework

The form of hybrid energy storage can effectively improve the economy of system operation, and the general framework of low-carbon economic dispatch of a new power system considering electric-hydrogen hybrid energy storage in this chapter is shown in Fig. 10.

4.4

Scheduling modeling example analysis

4.4.1

Introduction to the algorithm

In this paper, a modified IEEE30 node system is used for example analysis to verify the validity of the proposed model. The example contains a 1200 MW wind farm, a 400 MW photovoltaic plant, a 700 MW energy storage unit, and five thermal power units, and the relevant parameters of the energy storage and thermal power units are shown in Tables 2 and 3.

Table 2.

Storage parameter

Energy storage capacity/(MW·h)	700
Maximum charge/discharge power of energy storage/MW	120
Upper limit of state of charge	1
Lower limit of state of charge	0.3
Initial state of charge	0.6
Self-discharge rate/%	0.02
Charge/discharge efficiency/%	96

Table 3.

Engine parameters

Unit number	Lower limit of output	Upper limit of output	Cost parameters (a/b/c)/$ / (MW²·h)/$ / (MW·h)/($/h)	Start stop cost	Minimum start/stop time	Climbing power/(MW·h)
1	250	480	0.0464/30/196	2100	9/9	250
2	190	420	0.0510/30/187	1600	8/8	230
3	140	360	0.0118/18.7/162	1200	7/7	160
4	130	320	0.031/13.67/91.3	1150	5/5	110
5	80	160	0.286/16.92/84.5	950	4/4	80

A scheduling cycle of 24h and a unit scheduling period of 1h are specified. The penalty cost for wind and light abandonment is 160$/(MW-h) and the standby cost for thermal units is 30$/(MW-h), setting confidence level β₁ = β₂ = 0.95 in the standby opportunity constraint. The algorithm is computationally solved using the CPLEX solver. The wind and PV generation and load forecast curves are shown in Fig. 11.

4.4.2

Analysis of example results

In order to verify the effectiveness of considering wind power and photovoltaic power into system standby in high proportion new energy power system, the following two models are selected for comparative analysis.

Model 1: New energy sources participate in power system standby optimization.

Mode 2: new energy does not participate in power system backup optimization.

The results of the optimized scheduling scheme under different modes are shown in Table 4. Comparing the optimized scheduling results of Mode 1 and Mode 2, it can be seen that after the new energy is involved in the system standby optimization, the total cost of the system is obviously reduced, and although the standby cost of Mode 1 is slightly increased compared to Mode 2, the total cost of the system is reduced by a total of 6,383.67 U.S. dollars, as well as the actual power generation of the system’s wind power and photovoltaic power generation has risen instead by a total increase of 892.29 MW -h, indicating that the reasonable inclusion of new energy into the system standby can improve the level of new energy consumption, and overall improve the comprehensive benefits of the system.

Table 4.

Different patterns optimize the scheduling results

Mode	Total system cost/$	Backup cost/$	New energy generation capacity/(MW·h)
1	566028.14	286742.25	15267.63
2	572411.81	262593.69	14375.34

The actual generation power of new energy sources under different modes is shown in Fig. 12. As can be seen from Fig. 12, the actual generation power of wind power and PV in mode 1 program is more in the low load period, which is closer to the original new energy (including wind and PV) generation power prediction curve.

The results of the power output of each type of power supply in the power system under different modes are shown in Fig. 13, where (a)~(b) indicate the power output of mode 1 and mode 2, respectively.

As can be seen from Fig. 13, when new energy is involved in the system standby optimization in Mode 1, the overall power output of thermal power units is reduced, and the power generation cost of thermal power units is subsequently reduced, so the total cost of the system is lowered, and the economy of the scheduling scheme is improved. At the same time, the new energy consumption space of the system also increases, and the percentage of new energy generation in the system increases as shown in the calculation results.

The system standby optimization results under different modes are shown in Fig. 14, (a)~(b) represent the standby optimization results of mode 1 and mode 2 respectively.

As can be seen from Fig. 14(a), wind power and PV units tend to provide positive standby for the system during the load trough period, which is also originally in the period of high wind and light abandonment, in line with the expectation of the design of the scheduling program. Comparing Figure 14(a) and Figure 14(b), it can be seen that the new energy participation in system backup optimization can reduce the reserve capacity of conventional units, which can alleviate the system backup pressure to a certain extent. From the calculation results, it can be seen that when new energy is included in the system standby in a reasonable way, the output of thermal power units can be reduced to a certain extent when the load is in the low valley, reducing the reserved standby capacity of thermal power units, thus increasing the space for the consumption of new energy and reducing the amount of wind and light abandoned by the system.

5

Line loss visualization and monitoring analysis system design

On the basis of the prediction of power enterprise grid load and the construction of scheduling model, this paper designs a line loss visualization and monitoring analysis system to realize the line loss management and load prediction of power enterprise based on the optimization of new energy consumption technology.

5.1

Overall design

The overall objectives of the system design are as follows: 1)

Establish the data model of line loss visualization and monitoring and analysis system through information technology, realize the management of line loss basic data, complete the setting of data sources, provide interfaces for manual entry and automatic collection, calibrate the data quality, and ensure the authenticity, accuracy and reliability of basic data.

2)

Grid line loss information data visualization. Based on the geographic information of the power grid, it realizes the graphical display of the topological structure of power plants, substations, lines, switches, transformers, metering points and meters. Visualization of the association relationship of “station-line-transformer” file data in the form of a single-line diagram is realized, so as to achieve a unified display of the business systems at each source.

3)

Realize line loss data traceability. Trace back the line loss data from one layer to another, carry out data governance in the source system for the calculation data that cause abnormalities in the calculation of line loss, and monitor the process of transferring the line loss calculation data from the source system end to the calculation end of the line loss to ensure that the data transfer is successful.

4)

Realize panoramic display of line loss data. According to the needs of line loss refinement management, it displays the results of four-part line loss calculation of Hebei provincial power grid on the topological structure map, and real-time display according to sub-districts, sub-voltages, sub-stations, and sub-components, so as to quickly locate the areas of abnormal line loss, public transformers, and components.

5)

Realize monitoring and early warning of line loss data. The line loss indicators of each region and business unit are finely ranked, and the data quality of each source-side business system is finely evaluated, so as to realize the lean control of line loss work.

6)

Realize unified authority management to ensure the overall flexibility and security of the application.

5.2

Functional Module Design

The specific functional modules of the system mainly include three parts: panoramic display of power system line loss, line loss data visualization and line loss system monitoring and warning. 1)

Power system line loss panoramic display: display and analyze the line loss situation on SVG map according to voltage level, area, line and station, and show the files of station-line-transformer and gate meter graphically, and display the calculation results of power and line loss on the graph.

2)

Line loss data visualization: it includes indicator definition management, indicator monitoring management, basic indicator display and indicator traceability.

Indicator definition management: mainly defines the upper and lower thresholds of each indicator.

Indicator monitoring management: including supervising and rectifying, providing query and visualization of performance assessment indicators related to line loss rate, power supply, and power sales at provincial and municipal levels, and prompting rectification information for abnormal data monitored.

Basic indicator display: visualize the relationship of “station-line-transformer” and the profile of the gateway meter, and provide visual display of line loss rate, power supply and power sales related indicators.

Indicator traceability: Realize the traceability of line loss data, and carry out data governance in the source system for the calculation data that cause abnormalities in line loss calculation.

3)

Line loss system monitoring and early warning: including early warning rule setting, file variance monitoring desk: monitoring and early warning, distribution line monitoring and early warning, and other monitoring and early warning contents.

Early-warning rule setting: Based on the index evaluation standard of the four-point management index system of line loss, the early-warning rules for file variance, station area and distribution line are set according to the actual requirements of electric power enterprises.

File variation monitoring: monitor file variations in the line loss system, and provide alerts and warnings for file additions, updates and deletions that affect line loss modeling and power calculation.

Desk monitoring and early warning: according to the setting of desk early warning rules, it monitors and warns the indicators of power and line loss in the desk area.

Distribution line monitoring and warning: according to the warning rule settings for distribution lines, it monitors and warns on power consumption, line loss and other indicators of distribution lines.

5.3

Technical architecture design

Line loss visualization and monitoring analysis system based on J2EE specifications, using SSM framework to build MVC three-tier architecture, to achieve the B/S structure of the Web application mode, the use of HTML5+Canvas+JavaScript technology to display the foreground interface, the use of SVG technology to achieve the graphical visualization, and the choice of Oracle for the management of the system database.

The specific architecture design of the system is shown in Figure 15, which is hierarchically divided according to the layered architecture design method. 1)

Expression Layer

The interface of the system performance layer is mainly realized by using HTML5 tag library, JavaScript scripts and ECharts components, etc., and using HTTP protocol to realize the message transfer with the backend business logic layer of Web service and other services through the Ajax asynchronous refresh mechanism of the jQuery framework.

2)

Business logic layer

Business logic layer is the core part of the system design and implementation, covering the system business logic to achieve the essence of the algorithm, all business logic processing are realized in this layer. At the same time, it plays the role of a data channel, realizing the message transfer with the performance layer upwards and the data interaction with the persistence layer downwards.

In the online loss visualization and monitoring analysis system, the business logic layer uses SpringMVC to establish the connection between the performance layer and the business logic layer, to realize the response request from the front page, to verify the legitimacy of the data submitted by the front, and to realize the page switching work.

The business logic layer of the business implementation of this system uses the Spring framework to realize the Spring framework to establish the inversion of control (i.e., IOC) design pattern, provides a dependency injection and dependency lookup function, reasonably completes the object instantiation and configuration work, completes the functional encapsulation of the business logic and the interface reconfiguration, is the specific implementation of the business is transparent with respect to the upper level users.

3)

Persistence layer

The persistence layer is oriented to the system data layer and is responsible for completing the access operation to the database data. In this system, the persistence layer uses the MyBatis framework, MyBatis is an excellent lightweight framework, which supports data storage through JDBC. MyBatis implements a simple mapping relationship between entity classes and SQL, and Hibernate can only automatically generate SQL is different, in MyBatis will not be automatically generated at runtime. SQL execution, which provides a more flexible SQL statement mapping mechanism, you can manually adjust the parameters written to achieve the mapping relationship between the SQL statement and entity class JavaBean configuration, return to the target entity, this mechanism makes the persistence layer easier to design and maintain.

4)

Data layer

The data layer mainly includes the configuration data of the system connecting to the database data source and the relevant business data involved in the process of graphical presentation of line loss data and monitoring and analysis.

5.4

Database selection

The system in this paper adopts Oracle database for data management, which has high system compatibility and data portability, and has a big advantage in data management, can effectively manage spatial data, and can combine spatial data with attribute data, which can largely satisfy the system data storage and use requirements.

The process of data exchange in this system is as follows: the line loss visualization and monitoring analysis system collects real-time operation data and archive data from the integrated power and line loss management system; the spatial data of the power system stored in documents are manually imported into the database. The three functional modules of line loss visualization, line loss indicator monitoring and line loss indicator analysis in the system use SQL statements to interact with Oracle database. The data interaction between each module in the system and Oracle is shown in Figure 16.

6

Conclusion

In this paper, a grid load forecasting model based on LSTNet and federated learning and a load forecasting framework for sub-industries are constructed, and a new low-carbon and economic dispatch method for electric-hydrogen hybrid energy storage of the power system is proposed to realize the design of the line-lossable line-loss visualization and monitoring and analysis system. The conclusions are as follows:

First, the problem of high percentage of new energy consumption being hindered is analyzed. Due to the anti-peaking characteristics and random fluctuation characteristics of wind power, after large-scale wind power access to the grid, the peak-to-valley difference of the equivalent load of the system increases instead of decreasing, which increases the down-peaking demand of the system, and at the same time, the proportion of the conventional power supply access to the grid decreases, and the insufficient peaking capacity of the grid results in the obstruction of new energy consumption.

Second, with the increasing number of global aggregation rounds, the global model loss under different learning rates are decreasing, and in the three categories of industrial users, commercial users, and residential users, the global model all performs better in the load prediction at the learning rate of α = 1 × 10⁻⁴.

Again, comparing the model 1 in which new energy participates in the power system backup optimization and model 2 in which new energy does not participate in the power system backup optimization, it can be seen that after the new energy participates in the system backup optimization, the total cost of the system is significantly reduced, and the total cost of the system is reduced by a total of 6,383.67 USD in model 1 compared to model 2 although the backup cost is slightly increased and the actual power generation of the system for wind power and photovoltaic power generation is increased instead with an increase of 892.29 MW-per-minute. 892.29 MW-h, indicating that the reasonable inclusion of new energy into the system standby plays a role in improving the level of new energy consumption, and can improve the comprehensive benefits of the system on the whole.

Język:: Angielski

Częstotliwość wydawania:: 1 razy w roku
Dziedziny czasopisma:: Nauki biologiczne, Nauki biologiczne, inne, Matematyka, Matematyka stosowana, Matematyka ogólna, Fizyka, Fizyka, inne

Kanał RSS czasopisma

Research and Development of Line Loss Management and Load Forecasting System for Electric Power Enterprises Based on New Energy Consumption Technology Optimization

Jiadong Li

Shifang Hao

Shujun Ji

Kai Liu

Linhao Zhang

Data publikacji: 23 wrz 2025

Otrzymano: 17 sty 2025

Przyjęty: 20 kwi 2025

DOI: https://doi.org/10.2478/amns-2025-0970

Słowa kluczoweLSTNet, FedAvg, Federated learning, Load forecasting, New energy consumption, Line loss management

© 2025 Jiadong Li et al., published by Sciendo

This work is licensed under the Creative Commons Attribution 4.0 International License.

Słowa kluczowe
LSTNet, FedAvg, Federated learning, Load forecasting, New energy consumption, Line loss management