Big Data Algorithm for Resource Potential Awareness Response Optimization on the Power User Side Based on IoT Edge Computing

According to the agreed protocol, the IoT connects any object with the network through information-sensing equipment. The object exchanges and communicates information through the information dissemination medium to realize intelligent identification, positioning, tracking, supervision, and other functions [12, 13].

The typical architecture of the IoT is divided into three layers: perception layer, network layer, and application layer from bottom to top [14, 15]. The core capability of the perception layer to realize a comprehensive perception of the IoT is a part of the critical technology, standardization, and industrialization of the IoT that urgently needs breakthroughs. The key lies in having more accurate and comprehensive perception capabilities and solving low power consumption, miniaturization, and Low-cost issues. The network layer mainly takes the mobile communication network with extensive coverage as the infrastructure, the most standardized, industrialized, and mature part of the IoT. The key lies in optimizing and transforming the IoT application characteristics to form a system-aware network [16, 17]. The application layer provides rich applications, combines IoT technology with industry informatization needs, and realizes extensive intelligent application solutions. The key lies in industry integration, development, and utilization of information resources, low-cost and high- quality solutions, information security guarantees, and effective business models.

The IoT system is mainly composed of an operation support system, a sensor network system, a business application system, and a wireless communication network system [18]. The required information can be collected through the sensor network. In practice, customers can use RFID readers and related sensors to collect the required data. After the gateway terminal is aggregated, it can be smoothly transmitted to the designated application system through the wireless network.

The power IoT provides a variety of business IS for the construction of smart grid, including sensing detection, remote equipment monitoring, and equipment inspection [19]. These business requirements are diversified. For example, services such as video surveillance and equipment diagnosis have higher requirements for computing resources, while services such as smart meter monitoring and inspection robots are very sensitive to delays, so such services require more timely calculation results. At the same time, with the continuous construction and development of the power IoT, the data volume of power business terminals shows an explosive growth trend [20]. If only the traditional cloud computing mode were adopted, the pressure on the central cloud server and network transmission would increase sharply, resulting in the processing delay of most power tasks and making it difficult to meet business needs [21, 22]. In this context, edge computing technology is widely used to reduce the processing latency of power tasks.

Edge computing is a concept of nearby computing. Certain decisions place some computing operations on local computing devices closer to the data source so as to minimize the network cost and waiting time for data uploading to the cloud [23]. A core technology of edge computing is task allocation, which mainly includes two parts: task offloading and resource allocation [24]. Task offloading is the premise of resource allocation, which determines whether computing tasks need to be offloaded and which part of tasks should be offloaded. Resource allocation is the reasonable allocation of resources consumed by tasks in the process of offloading. The main difference between the two is that the former belongs to the planning stage, while the latter belongs to the implementation stage. The so-called edge computing task offloading refers to taking advantage of edge computing to place some tasks that should have been uploaded to the central cloud server for execution on the edge server closer to the user side for execution, thereby achieving lower task transmission delay [25]. More specifically, edge computing task offloading refers to the process of reasonably allocating tasks to edge devices with strong computing power, such as edge servers, under the condition of meeting the specified constraints [26]. These constraints usually include factors such as network communication service quality and load balancing of edge servers.

With the continuous development of computer technology, perception technology is also constantly evolving, and finally a dynamic decision model based on Endsley perception model is formed [27]. Endsley dynamic decision model mainly divides perception into three levels, and combines system factors and individual factors to assist the perception system in making intelligent decision-making and intelligent control functions. The schematic diagram of the Endsley dynamic decision model is shown in Figure 1.

Figure 1.

Schematic diagram of dynamic decision model

Element awareness is the first level of perception, that is, collecting various information through perception means to form a preliminary understanding of the outside world, that is, the stage of data collection and data preprocessing. The data acquisition stage collects information through various sensors, remote sensors, and other technologies, and the data preprocessing stage carries out preliminary processing on the collected information, including smooth noise, filling, deleting integration, protocol, transformation, and other steps.

Understanding is the key part of perception, that is, in-depth analysis of the collected information to form a systematic understanding of the outside world. This step mainly includes feature extraction, classification recognition, pattern recognition, and other steps [28]. In the feature extraction stage, the key information in the data is extracted to facilitate the subsequent classification recognition and pattern recognition. In the classification recognition stage, the data is divided into different categories according to the extracted features to facilitate further pattern recognition. In the pattern recognition stage, the data of each category is deeply analyzed to identify the external patterns so as to facilitate the real-time grasp of the outside world.

Prediction is the ultimate goal of perception, that is, to predict future development according to the mastery of the outside world. This step includes model establishment, simulation, and other steps. Among them, the model establishment stage establishes a prediction model according to the existing external model, while the simulation stage simulates future development according to the established model so as to facilitate the prediction of the external world. It is also necessary to comprehensively assist decision-making and execution with the help of system factors such as automation and system capabilities and individual factors such as personal experience and personal capabilities. Endsley Dynamic Decision Model is a complex system that realizes a comprehensive understanding of the outside world through three aspects: awareness, understanding, and prediction. The progress of perception technology not only helps to improve the efficiency of decision-making but also has great significance for solving various security threats and environmental problems.

The demand-side response has very important social benefits. In the power market competition, power users will change their power consumption mode according to their own power consumption situation and real-time electricity price adjustment so as to maximize their benefits while meeting their power demand. This will reduce the power load during peak power consumption periods to a certain extent, thereby reducing the pressure on the power generation side. At the same time, it can also reduce the real-time electricity price in the market [29, 30]. In addition, transferring part of electricity consumption during peak periods to low periods can also improve the overall power generation efficiency of power plants. It can be seen that through the interaction between users and power plants, the benefits of system demand-side response can be maximized, and the efficiency of the power generation side can be improved.

On the power generation side of the power system, whether the unit is started or stopped must comprehensively consider various factors such as power generation costs and users’ electricity consumption. The power generation costs change nonlinearly, and users’ electricity consumption is also constantly changing. Therefore, different decision results directly determine different real-time electricity prices in the market. Therefore, aiming at the current start-stop problem of generating units, this study establishes a corresponding power model under the condition of considering the change of power load on the demand side. The model uses the Lagrange algorithm to calculate the results. In this model, compound bidding is adopted, and the power generation decision considers the impact of unit start-stop on the real-time electricity price in the market. According to the incentive effect of real-time electricity prices, power users actively change the electricity consumption situation and transfer unnecessary electricity consumption during the peak period to the low period. Based on this result, the cost and benefit change indicators of the power system, the power consumption side, and the power generation side before and after the demand response is established and quantitatively assessed, and the start-stop changes of units after the demand-side response and the corresponding electricity price fluctuations are considered from two aspects: comprehensive benefits and marginal benefits; The impact and changes on the benefits of the power generation side and the power consumption side after the start-up and stop of the unit are analyzed, so as to evaluate the social benefits brought by the demand-side response.

Demand-side response requires the electricity consumption side to actively change its electricity consumption behavior based on the incentive effect of real-time electricity prices and transfer part of the electricity during the peak period of electricity consumption to the low period of electricity consumption, thereby reducing the power generation burden on the power generation side and reducing the market electricity price. Unstable and sudden changes reduce market risks and achieve efficient operation of the power system. Under normal circumstances, demand-side response is generally carried out in the day-ahead market to ensure that users can adjust their electricity consumption behavior.

In order to quantitatively evaluate the market benefits, the evaluation tool selected in this paper is the power trading model of the competitive power bank. The power bank market now generally adopts 24 or 48 trading hours a day, that is, every hour or half an hour for transactions. This article adopts the trading mode of 24 trading hours per day. The power generation bidding adopts the compound bidding mode; that is, it includes the start-up cost of the unit, the cost of increasing power generation, and the cost of power generation without load. Electricity-side bidding indicates the change in electricity consumption with real-time electricity prices. The market design takes into account the demand transfer on the electricity side, and users’ electricity demand can shift from peak periods to trough periods. The ultimate goal of the transaction is to minimize the cost of electricity purchase under the bids given by the generation side. The electricity market in this paper adopts a perfect competition market, and it is considered that the bid given by the power generation side is its marginal power generation cost, so the goal of minimizing power purchase cost can be converted into minimizing power generation cost, that is, as shown in formula (1): (1) f(x)=min∑t=1TOCt+ui,t+FCi+MCi,k

In formula (1), OC^t represents the power generation cost of the system in time period t, u^i,t represents the start-stop state of unit i in time period t, FCⁱ represents the cost of unit power generation without load, and MC^ik represents the marginal cost of piecewise linear cost of unit i in time period k. The cost curve of the generator set is generally expressed by a quadratic curve as shown in Equation (2), where t is the shutdown time, and P^i,k,t represents the output of the k-th section of the piecewise linear cost of the unit i in the period t. S^i,t represents the start-up cost of unit i in time period t and τ represents the cooling time constant of boiler.

The three marginal costs expressed in the piecewise linear cost curve of unit i can be expressed as equations (3)-(5) respectively, aⁱ and bⁱ are the turbine start-up cost and boiler start-up cost respectively when the unit is started. P_min and P_max are the minimum and maximum outputs of the unit respectively, and e₁, and e₂ are the two segment points of the curve respectively.

This study proposes a suite of edge IoT data processing platform software architecture to support IoT real-time data access, data forwarding, big data processing, data storage, and online machine learning applications. The platform deploys distributed application services at each layer based on container technology and opens up the data inflow and outflow between applications at each layer through the container orchestration capabilities and load-balancing services provided by Kubernetes. As shown in Figure 2, the system architecture consists of an IoT data access layer, a data forwarding cache layer, a big data processing layer, an online machine learning application layer, and a data storage layer, and provides data distribution among distributed component instances of each layer through the load balancing Service of Kubernetes. Specifically, the IoT Data Access Layer uses distributed IoT access components, is containerized and deployed on lightweight edge computing nodes such as ARM architecture servers, and at the same time provides message parsing services and data rule engines for protocol data such as MQTT, directly interfaces with data flows reported by IoT terminal devices, and supports data forwarding services with various distributed message middleware. The Data Transfer Layer mainly deploys distributed message middleware Kafka and ZooKeeper clusters to provide caching services for data forwarded by the IoT data access layer, realizing the original data access and big data of the IoT. The decoupling of the computing engine improves the stability of the platform system. The Big Data Processing Layer deploys Flink, a stream computing data processing engine, to provide big data processing functions for real-time data of the IoT, including stream computing, sliding window computing, and batch processing, and then stores the real-time computing results in the downstream database.

Figure 2.

Architecture diagram of edge IoT data processing platform

The online machine learning application layer (Online ML layer) uses the Tensorflow.js framework to provide a pipeline of lightweight online machine learning applications, including online training (Real-time Model Training), online verification (Real-time validation) and Online Prediction, three parallel processes; The online training process periodically pulls the latest big data stream calculation results as the training set and obtains the latest model through training. In the online verification part, the newly generated model is loaded into memory, and the model is verified by using part of the data set. In the online prediction process, the latest model is also applied to the real-time online data stream, and the inference results output by the model are obtained, and the results are written into the time series database. The Data Storage Layer is built on the edge data storage server, uses the time series database to receive the result data of big data stream processing and online machine learning applications, and also serves as the middleware of big data processing and machine learning applications. This layer also provides a visual display interface of application monitoring and stream processing results.

The scheduling algorithm needs to consider factors such as task execution time period, so the scheduling algorithm uses the training Start time, training End time, and resource state of edge nodes of edge machine learning training tasks as the input state of the DQN neural network, and uses S = {Start, End, NodeRes} to represent the input state space set. Among them, Start is a matrix of M × N dimensions, which represents the set of each training Start time of M Inference tasks on the edge computing platform starting from the timestamp of each round of scheduling process, corresponding to 1-N The difference between the training Start time and the scheduling process Start time. For example, if the start time of the scheduling process is 0 seconds, and one machine learning task completes three training tasks in the scheduling period, and the difference between its start timestamp and the start time of the scheduling process is 2 seconds, 15 seconds and 25 seconds respectively, then the start time set of this task is {2, 15, 25}. Similarly, End represents a matrix of M × N dimensions, which means the set of 1-N training End times of M Inference tasks on the edge computing platform. NodeRes is an M × K-dimensional matrix, which means the running distribution state of M Inference tasks running on K edge nodes in a scheduling time period.

The scheduling action space output by the DQN scheduling algorithm is a matrix of M × K dimensions, which represents the probability Q value of scheduling M machine learning tasks to K edge computing nodes. The greater the probability Q value, the greater the action trend of the DQN neural network to schedule tasks to the node in this scheduling process, as shown in the following equation (6): (6) Qπ(s,a)=(Q11⋯Q1|K|⋮⋱⋮QM1⋯QM||K∣)

The agent takes the maximum probability Q value in the matrix each time to obtain its corresponding position coordinates in the probability matrix, which are used as the machine learning task m and the target node k to be scheduled, as shown in (7) below: (7) (m,k)=argmax∀at∈AQπ(s,a) argmax represents the set of argument values when a function takes the maximum value in its domain. In this paper, the reward calculation is based on the average Job Completion Time (JCT) of online machine learning tasks on the edge computing platform. The scheduling process needs to output multi-step scheduling decision actions for multiple Inferences, and then interact with the environment together with all Inference scheduling decisions. After the scheduling is completed, the reward calculation is performed based on the observed average JCT at the beginning of the next round of scheduling process. In other words, in order to ensure that each Inference obtains a scheduling decision, it is necessary to control the DQN neural network to perform M continuous and non- repetitive action outputs to obtain the target edge node of each machine learning task to be scheduled. Then, each task is actually scheduled and orchestrated. Among them, the reward obtained by the j-th output action of the i-th training episode is defined as equation (8): (8) reward(i,j)=γ(M−j) reward i, j∈1,…,M Because the reward given by the environment can only be obtained after the agent interacts with the environment and performs scheduling when the DQN network outputs the last scheduling decision. Therefore, this paper uses γ as the reward attenuation factor, and the closer the reward obtained by the M-th action, the greater the correlation with the final reward obtained by this iteration. The reward obtained by the i-th episode is defined as equation (9): (9) reward i={ 0,i=1;rw+λ1,i≥2,cost(si,M)≥0,MinCost>cost(si,M);rw,i≥2,cost(si,M)≥0,‖ MinCost−cost(si,M) ‖<δ;λ2·(MinCost−cost(si,M)), otherwise. Among them, λ₁ and A₂ are variable reward control parameters. In this paper, the value of λ₁ is set to 100 and the value of X2 is 10. rw is used to adjust the current reward. When the minimum cost value in history is greater than the cost (s_i, m) obtained in this iteration, rw automatically increases λ₁ value. The Cost value obtained by the i-th episode output scheduling decision can be calculated by the following formula (10): (10) cost(si,M)=λ3(ω1JCT1+…+ωKJCTK)+λ4Var(Numi) Where Num_i represents the i-th episode, the number matrix of tasks running on K edge nodes. Solve the variance for the number of node tasks matrix as part of Cost. ωk denotes the weight of the average completion time of online machine learning tasks running on the k-th edge compute. λ₃ and λ₄ are also variable weight coefficients. In this paper, the value of λ₃ is set to 0.01 and the value of λ₄ is set to 20. Since offline simulation also needs to estimate the average JCT of each task to get the reward value in the current state. According to the foregoing, the completion time of the task is related to the computing resources actually used by the task running on the current node. When a plurality of machine learning tasks is run on a node, the computing resource usage time slice obtained by each task is relatively reduced. According to the experimental results of comparing the average JCT and execution overlap time of online machine learning tasks shown, it is assumed that the JCT of the task running on edge node k is related to the overlap time interval between the average JCT of the task and the task running on node k, and it is estimated by the following calculation method, as shown in Equation (11): (11) JCTm(k)=JCTmå(k)+λ5N(k)+λ6Num(k) $$JC{T_m}(k) = JCT_m^{a}(k) + {\lambda _5}N(k) + {\lambda _6}Num(k)$$ Among them, N (k) represents the sum of task overlapping times on the k-th edge node, which is calculated by the start time and deadline time of all online machine learning tasks on the node. Num (k) denotes the number of tasks on the k-th edge node. JCT^*m(k) represents the average JCT when only the m-th machine learning task is running on edge node k, which can be actually measured by deploying the task to the node. λ₅ and λ₆ are likewise variable weight coefficients.