Application of knowledge graph in smart grid fault diagnosis

The fault diagnosis of power equipment is special; so, the construction of its knowledge graph requires a series of modifications.

Fig. 1

When the power grid fails, hundreds of accidents, displacement or abnormal signals will be generated in a short period of time so that the staff cannot comprehensively analyse the fault alarm information, resulting in low efficiency and unsatisfactory information judgement. Therefore, for the knowledge map combined with language processing technology, the knowledge of fault alarm information is automatically parsed to form a structured representation that can be understood and calculated by machines, and then, query matching and judgement analysis are performed based on the device information and business knowledge stored in the knowledge graph. The technical process of fault information analysis is shown in Figure 2.

Fig. 2

When the power grid system fails, it is necessary to use the comprehensive intelligent alarm and other information as the trigger signal and use the fault information capture technology to ensure the complete acquisition of remote signalling information when the fault occurs and to eliminate abnormal alarm signals unrelated to the accident; the obtained fault information is analysed one by one through the information extraction technology, and the logical variables required for fault diagnosis are obtained [17]; information matching is to quickly match the fault alarm information after structural analysis from the equipment entity map of the power grid to obtain the corresponding network topology and electrical equipment and meteorological and environmental data in the power loss area; knowledge query It is linked to the relevant knowledge of scheduling procedures, rules and emergency plans for logical analysis of actions and fault diagnosis. The graph data structure of the knowledge graph often has faster query efficiency and can meet the speed requirements of fault diagnosis; according to the compatibility relationship between equipment, circuit breakers and protection, as well as causal relationship, mutual exclusion relationship and control logic, signal error correction is realised And filtering, to determine the authenticity or lack of information, and then complete the action logic judgement and fault information correction, to achieve power grid fault signal analysis and fault mode judgement [18].

After completing the information structuring and logical judgement, the key information flow of the current grid failure can be recorded in the form of triples of the graph, and the important abnormal or fault information can be linked to the relevant analysis or disposal knowledge nodes in the business logic graph so as to realise the complete expression of fault explicit and tacit knowledge [19].

The power grid fault diagnosis system is a system that uses directed arcs to connect the two groups of nodes of the attribute point (circular node) and the transition (bar node). The number of instructions on the attribute point changes when the event occurs [20]. Dependent points, transitions and directed arcs, respectively, construct process states, occurrence events and process evolution laws. As shown in Figure 3, the belonging point p₁ is not only the input and output belonging point of the transition points 1 but also the input belonging point of the transition s₂. When transition s₂ fires, it removes one instruction from point p₁ and puts two instructions into point p₂, and the instruction at p₂ can then be used for transitions s₃ and s₄.

Fig. 3

Since the model will have different evolutions in each state, the dynamic behaviour of the model should be studied, and the distribution of the instructions on the attribute points is called the model mark, and the current state of the model system is defined by this [21].

A model token is usually represented as follows: (1) Q=(P,S,P−,P+) Q = \left({P,S,{P^ -},{P^ +}} \right) where P is the set of n belonging points, namely: (2) P={p1,p2,…,pn} P = \left\{{{p_1},{p_2},\ldots{},{p_n}} \right\}

S is the transition set of m, then: (3) S={s1,s2,…,sn} S = \left\{{{s_1},{s_2},\ldots{},{s_n}} \right\}

P⁻: P×T → N and P⁺ : P×S → N are the matrices of the weights of directed arcs from the specified belonging point to the transition and the weights of the directed arcs from the transition to the belonging point, respectively, where N is a set of non-negative numbers and 0 means no directed arc. In particular, P⁻(i, j) is the weight of the directed arc from the belonging point p_i to the transition s_j, and P⁺ (i, j) is the weight of the directed arc from the transition s_j to the belonging point p_i. Then, the evolution equation of Q is written as follows: (4) S[k+1]=S[k]+(P++P−)⋅σ[k]=S[k]+M⋅σ[k] S\left[{k + 1} \right] = S\left[k \right] + \left({{P^ +} + {P^ -}} \right)\cdot \sigma \left[k \right] = S\left[k \right] + M \cdot \sigma \left[k \right]

In the formula, M ≡ P^± −P⁼ is an order correlation matrix n × m, S [k] is the identification vector of Q at each time node k, which is of order n × 1, and σ [k] is the trigger vector of order n × 1, which is an indicator vector, indicating the transition triggered at time k [22].

If and only if the instruction between each input attribute point p_i to transition is greater s_j than or equal to the directed arc weight between p_i and s_j, the transition s_j ∈ S is enabled at time k; it can be described as S [k] ≥ P⁼ (:, s_j), where P⁼ (:, s_j) represents the P⁼ column of s_j. P^± and P⁼ are written as follows: (5) P−=[11000001100000100001]; P+=[10001020000010000010] {P^ -} = \left[\begin{matrix}1 & 1 & 0 & 0 & 0\\0 & 0 & 1 & 1 & 0\\0 & 0 & 0 & 0 & 1\\0 & 0 & 0 & 0 & 1\end{matrix} \right];\quad {P^ +} = \left[\begin{matrix}1 & 0 & 0 & 0 & 1\\0 & 2 & 0 & 0 & 0\\0 & 0 & 1 & 0 & 0\\0 & 0 & 0 & 1 & 0\end{matrix} \right]

In the above formula, the columns and rows represent transitions and attribute points, respectively. The correlation matrix can be expressed as follows: (6) M≡P±−P==(0−100102−1−100010−10001−1) M \equiv {P^{\pm}} - {P^{=}} = \left(\begin{matrix}0 & - 1 & 0 & 0 & 1\\0 & 2 & - 1 & -1 & 0\\0 & 0 & 1 & 0 & - 1\\0 & 0 & 0 & 1 & - 1\end{matrix} \right)

At time k = 0, there are written as follows: (7) S[1]=[1000]T S\left[1 \right] = \left[\begin{matrix}1 & 0 & 0 & 0\end{matrix} \right]^{T}

Triggering of transition s₂ results in the following equation: (8) S[1]=S[0]+M×[0100]T=[0200]T S\left[1 \right] = S\left[0 \right] + M \times \left[\begin{matrix}0 & 1 & 0 & 0\end{matrix} \right]^{T} =\left[\begin{matrix}0 & 2 & 0 & 0\end{matrix} \right]^{T}

From the above, the triggering of transition s₂ can indicate the occurrence of an event (such as a circuit breaker tripping), causing the system to transition to another state (such as an open circuit breaker). A fault in the power system can cause the command quantity in the home point to become negative [23].

To identify system failures, construct a redundant pseudocode Q_H. Redundancy Q_H is designed by modulo operation based on linear error-correcting codes over finite fields GF (p).

Suppose formula (4) is introduced into the system and d attribute points are added to the original system, namely: (9) SH[k]=[In−C−*]S−[k] {S_H}\left[k \right] = \left[{\frac{{\mathop {{I_n}}\limits_{-}}}{{{{\mathop C\limits_{-}}^{*}}}}} \right]\mathop S\limits_{-} \left[k \right]

All k and I_n in Eq. (9) are a matrix of order n × n, and C^* is a matrix (d×n) of non-negative integers of order.

Therefore, the evolution of the redundant system can be parameterised as follows: (10) SH[k+1]=SH[k]+[P+−C−*P+−−D−]σ−[k]−[P−−C−*P−−−D−]σ−[k]=SH[k]+Γ−+σ−[k]−Γ−−σ−[k] {S_H}\left[{k + 1} \right] = {S_H}\left[k \right] + \left[{\frac{{\mathop{{P^ +}}\limits_{-}}}{{{{\mathop C\limits_{-}}^{*}}\mathop {{P^ +}}\limits_{-} - \mathop D\limits_{-}}}} \right]\mathop \sigma \limits_{-} \left[k \right] -\left[{\frac{{\mathop {{P^ -}}\limits_{-}}}{{{{\mathop C\limits_{-}}^{*}}\mathop {{P^ -}}\limits_{-} - \mathop D\limits_{-}}}} \right]\mathop \sigma\limits_{-} \left[k \right] = {S_H}\left[k \right] + {\mathop \Gamma \limits_{-}^{+}}\mathop \sigma \limits_{-} \left[k \right] - {\mathop \Gamma \limits_{-}^{-}}\mathop \sigma \limits_{-} \left[k \right]

In formula (10), D is a non-negative integer matrix d × n of order, and d additional attribute points and n original attribute points constitute an embedded Q_H redundant system. Assuming that the interval of a period of time is t, then it can be written as follows: (11) t=[1,2,…,K] t = \left[{1,2,\ldots{},K} \right]

Identify the home point and transition fault in this time interval, assuming that e+t− \mathop {{e^{+}}_t}\limits_{-} , e−t− \mathop {{e^ -}_t}\limits_{-} and e_P represent the indicator vectors of the post-transition fault, the pre-transition fault and the home point fault, respectively; the fault system state at time K can be expressed as follows: (12) Sf−[k]=S−[k]−Γ−+×e+t−+Γ−−e−t−+eP− \mathop {{S_f}}\limits_{-} \left[k \right] = \mathop S\limits_{-} \left[k\right] - {\mathop \Gamma \limits_{-}^{+}} \times \mathop {{e^ +}_t}\limits_{-} + {\mathop \Gamma \limits_{-}^{-}}\mathop {{e^ -}_t}\limits_{-} + \mathop {{e_P}}\limits_{-}

If the diagnosis of power grid fault is to be successfully completed, a concurrent vector needs to be integrated and the evaluation of this quantity should be completed. At time K, the concurrency vector is defined as follows: (13) s−[K]=FSf−[k] \mathop s\limits_{-} \left[K \right] = F\mathop {{S_f}}\limits_{-} \left[k\right]

The above formula (13) also satisfies the following formula: (14) s−[K]=D−(e+t+−e−t−)+[−C−*Id−]eP− \mathop s\limits_{-} \left[K \right] = \mathop D\limits_{-} \left({\mathop{{e^ +}_t +}\limits_{-} \mathop {{e^ -}_t}\limits_{-}} \right) + \left[{-{{\mathop C\limits_{-}}^{*}}\mathop {{I_d}}\limits_{-}} \right]\mathop{{e_P}}\limits_{-}

If no fault occurs, e+t− \mathop {{e^ +}_t}\limits_{-} , e−t− \mathop {{e^ -}_t}\limits_{-} and e_P are all zero.

First, the fault at time K must be obtained. Assuming that the fault mode of the concurrent point is p and the evaluation of p is completed, the concurrent fault can be expressed as follows: (15) sP−[K]=[−C−*Id−]eP−(modp) \mathop {{s_P}}\limits_{-} \left[K \right] = \left[{- {{\mathop C\limits_{-}}^{*}}\mathop {{I_d}}\limits_{-}} \right]\mathop {{e_P}}\limits_{-} \left({\bmod p} \right)

When no fault is found at the belonging point or the faults of all the belonging points are acquired, then the concurrent transition fault at time K can be expressed as follows: (16) st−[K]=D−(e+t−−e−t−) \mathop {{s_t}}\limits_{-} \left[K \right] = \mathop D\limits_{-} \left({\mathop {{e^ +}_t -}\limits_{-} \mathop {{e^ -}_t}\limits_{-}} \right)

s_t [K] can be seen from the above that it is impossible to determine multiple faults, and the common influencing factors of the transition of the belonging point are pre-faults and post-faults. Therefore, the structures of the matrices D and C determine the number of faults finally determined. Suppose x_t is the input word at time t, then E_t and E_t^′ represent the fault state and temporary fault state, respectively, and h_t is the hidden layer state of the belonging point. By forgetting and memorising the state information of attribute points, the system can have memory for useful fault information and forgetfulness for invalid fault information [24]. f_t, i_t and o_t represent the forget gate, the input gate and the output gate, respectively. When the output value is between 0 and 1, the input quantity φ represents the number that can be released. 0 means that all are not allowed to be released, and 1 means that all are allowed to be released; then, the number of releases can be expressed as follows: (17) φ(z)=11+e−z \varphi \left(z \right) = \frac{1}{{1 + {e^{- z}}}}

Assuming that the state of the hidden layer of the attribute point at the last moment is h_t−1, the state of each attribute point of the output value between 0 and 1 is E_t−1; at this time, the output value determines the number of the attribute point state at the previous moment retained so far; W_f and b_f respectively represent forgetting weight matrix and forgetting gate bias term; the number of forgetting gates can be expressed as follows: (18) ft=φ(Wf⋅[ht−1,xt]+bf) {f_t} = \varphi \left({{W_f} \cdot \left[{{h_{t - 1}},{x_t}} \right] + {b_f}}\right)

Assuming that the output value between 0 and 1 determines the number of the current state of the input point that has been retained so far, W_i and b_i represent the input gate weight matrix and the input gate bias term, respectively; the number of input gates can be expressed as follows: (19) it=φ(Wi⋅[ht−1,xt]+bi) {i_t} = \varphi \left({{W_i} \cdot \left[{{h_{t - 1}},{x_t}} \right] + {b_i}}\right)

The status of the belonging point can be updated by combining the above conditions of the input gate: (20) Et'=tanh(WE⋅[ht−1,xt]+bE) {E_t}^{\prime} = \tanh \left({{W_E} \cdot \left[{{h_{t - 1}},{x_t}} \right] + {b_E}}\right) (21) Et=ft⋅Et−1+it⋅Et' {E_t} = {f_t} \cdot {E_{t - 1}} + {i_t} \cdot {E_t}^{\prime}

The current state of the belonging point has been retained until now, and the output values are represented by h_t, W_o and b_o, then the output gate formula can be expressed as follows: (22) ot=φ(Wo⋅[ht−1,xt]+bo) {o_t} = \varphi \left({{W_o} \cdot \left[{{h_{t - 1}},{x_t}} \right] + {b_o}}\right)

The current output value can be expressed as follows: (23) ht=ot⋅tanh(Et) {h_t} = {o_t} \cdot \tanh \left({{E_t}} \right)

Taking the power grid company of a certain city as the experimental data source, the power grid company's power grid site fault quantity, fault location and fault analysis information are obtained, and the data volume reached 5.3TB. In the process of building a graph, structured data are guided by fields, semi-structured data are guided by data indicators extracted from fields and unstructured data are guided by word segmentation.

Based on the internal equipment fault alarm signal of the pilot power grid company, the system proposed in this paper is used to analyse the alarm signal, which helps the monitoring personnel to quickly lock the equipment mechanism or abnormal parameter that may have problems, and obtain the following abnormal signals of the power grid equipment and the signal generating signal. Possible reasons are as in Table 1.

As it can be seen from Table 1, for the abnormal alarm signals issued by the power grid equipment, the knowledge graph fault diagnosis system proposed in this paper can analyse the causes of these abnormal alarm signals and provide the staff with the reasons for these abnormal alarm signals through the cause analysis and the location of the fault to achieve rapid diagnosis of power grid equipment faults.