Cable core temperature dynamic early warning method based on multi-source parameter inversion

Abstract

The invention discloses a multi-source parameter inversion-based cable core body temperature dynamic early warning method, which comprises the following steps of: inputting parameters such as cable surface temperature, air convection coefficient and thermal resistance under an actual working condition into an established cable core temperature inversion model to obtain a cable core temperature predicted at the current moment; determining a temperature change threshold value during normal work through the mean value and the standard deviation of the temperature change values during normal work of the cable core, and superposing the temperature change threshold value with a cable core temperature predicted value at the previous moment to form a cable core temperature dynamic early warning threshold value at the current moment; if the cable core temperature at the current moment exceeds the dynamic early warning threshold value, it is considered that the cable state may be abnormal. According to the invention, the cable core temperature is calculated by using a cable core temperature inversion model trained by using a loss function introducing weight adjustment and a regularization item through the surface temperature and thermal resistance of the cable, utilizing the current-carrying capacity and the heat conduction characteristic of a material and considering the loss of grounding circulation, and the cable core temperature is compared with a cable core temperature early warning threshold value; and dynamic early warning of the cable temperature is realized.

Description

Technical Field

[0001] The invention relates to the field of electric power technology, and in particular to a cable core temperature dynamic early warning method based on multi-source parameter inversion. Background Art

[0002] With the continuous expansion of urban scale and the increase of electricity load, cable lines have gradually become the main power supply mode of urban power grids. Among them, cross-linked polyethylene insulated power cables account for more than 80% of the cables in operation. When the temperature inside the cable is too high for a long time, it will accelerate the aging of the insulation material and reduce the service life of the cable. Therefore, the cross-linked polyethylene cable limits the maximum allowable temperature of the conductor to no more than 90°C. At present, for the measurement of cable line temperature, distributed temperature sensing optical fiber is mainly used, and the cable surface laying method is adopted. Under this laying method, the temperature measuring optical fiber can realize distributed and long-distance monitoring of the cable surface temperature, but the internal temperature of the conductor cannot be obtained.

[0003] In the existing technology, the cable temperature prediction model has limitations when processing complex data with strong time dependence, and cannot effectively capture long-term dynamic trends and short-term rapid fluctuations. The proposed cable core temperature inversion model can more accurately understand the dynamic temperature changes of the cable through a double-layer structure: the first layer of LSTM is used to capture the temperature change characteristics in a short period of time (such as current carrying capacity fluctuations in a short period of time). The second layer of LSTM is used to capture the temperature change trend over a longer time range (such as the impact of long-term changes in thermal conductivity with ambient temperature). And the loss function that introduces weight adjustment and regularization terms is used for training, which can effectively reduce the sensitivity to outliers during calculation and improve the generalization ability of the model.

[0004] On the other hand, in the existing technology, the model often only considers the surface temperature and current carrying capacity of the cable, while ignoring other factors that may affect the core temperature. The input parameters of the proposed method include cable surface temperature, current carrying capacity, material thermal conductivity and ground loop loss. Through the combination of multi-source data, the multi-variable input layer (including surface temperature, current carrying capacity, material thermal conductivity and ground loop loss) can help the model more comprehensively describe the working state of the cable, especially considering the influence of thermal conductivity and energy loss. The model can learn the relationship and effect between different variables and capture more complex influence patterns.

[0005] The cable monitoring system in the prior art generally uses a static temperature threshold value, and issues an alarm when the cable core temperature exceeds a preset upper temperature limit. This static threshold value does not take into account the complexity of actual working conditions and is prone to false alarms or missed alarms. Summary of the invention

[0006] In view of the problem that the cable conductor temperature is difficult to measure directly, the present invention provides a cable core temperature dynamic warning method based on multi-source parameter inversion. The temperature change threshold during normal operation is established through the mean and standard deviation of the temperature change value during normal operation of the cable core, and is superimposed with the cable core temperature prediction value at the previous moment to form the cable core temperature dynamic warning threshold at the current moment. When the cable core temperature exceeds the dynamic threshold, the model triggers an action warning. This dynamic threshold can avoid the limitations brought by the static threshold setting and reduce the probability of false alarms or missed alarms. This method will help to more accurately understand the actual operating conditions of the cable.

[0007] The present invention discloses a cable core temperature dynamic early warning method based on multi-source parameter inversion, which comprises:

[0008] The surface temperature, air convection coefficient, thermal resistance, average ambient temperature, average operating temperature, current carrying capacity, ground loop current loss and skin coefficient of the cable under actual working conditions are input into the long short-term memory neural network model based on the multi-head attention mechanism to obtain the predicted cable core temperature at the current moment;

[0009] The temperature change threshold during normal operation is established through the mean and standard deviation of the temperature change value of the cable core during normal operation, and is superimposed with the cable core temperature prediction value at the previous moment to form the dynamic warning threshold of the cable core temperature at the current moment; if the cable core temperature at the current moment exceeds the dynamic warning threshold, it is considered that the cable status may be abnormal.

[0010] Furthermore, the long short-term memory neural network model is composed of a sequence input layer, a first LSTM layer, an activation function layer, a Dropout layer, a second LSTM layer, a multi-head attention mechanism layer, a fully connected layer and a regression layer connected in sequence. A loss function is used during training, and weight adjustment and regularization terms are introduced to improve the generalization ability of the model;

[0011] The sequence input layer is used to input the air convection coefficient α, surface temperature T, and average operating temperature T of the cable. W , average ambient temperature T E The time series X at the current time t is composed of the current carrying capacity I, the real-time skin coefficient ρ, the thermal resistance R and the ground loop loss P t ;

[0012]

[0013] The first LSTM layer includes a memory unit, which includes an input gate, a forget gate, and an output gate. The input gate is used to control the new information added to the memory unit at the current moment. The output of the input gate is used to determine the new information added to the memory unit. The input gate generates a candidate memory unit, which represents the new information content that can be introduced at the current moment, and generates a new candidate state. The forget gate is used to determine the information that needs to be retained and discarded in the memory unit at the previous moment. The output gate is used to determine the output of the hidden state at the current moment.

[0014] The Relu activation function of the activation function layer is used to optimize the hidden state output by the first LSTM layer for training; the Dropout layer is used to reduce the overfitting effect of the hidden state after optimization training and output neurons; the second LSTM layer is used to receive the neurons output by the Dropout layer and output the hidden state; the multi-head attention mechanism layer is used to receive the hidden state output by the second LSTM layer, and use weighted splicing of the hidden state output by the second LSTM layer and the multi-head attention mechanism, and output the splicing result; the fully connected layer is used to receive the splicing result and output the predicted cable core temperature.

[0015] Furthermore, the input gate uses the following formula to calculate the received time series X t ={X1,X2,…,X t} for processing:

[0016]

[0017] in, is the output of the input gate at time t, h t-1 is the hidden state at the previous time t-1, C i is the weight matrix, z i is the bias term, σ is the sigmoid function; the input gate generates a candidate memory unit, which represents the new information content that can be introduced at the current moment, and generates a new candidate state K t ;

[0018] K t =tanh(C K ·[h t-1 ,X t ]+z K )

[0019] Among them, tanh is the hyperbolic tangent function, C K is the weight matrix, z K is the bias term.

[0020] Furthermore, the forget gate uses the following formula to calculate the received time series X t ={X1,X2,…,X t} for processing:

[0021] f t =σ(C f ·[h t-1 ,X t ]+z f )

[0022] Among them, f t is the information that needs to be retained by the forget gate, σ is the sigmoid function; [h t-1 ,X t ] is the hidden state h at the previous time t-1 t-1 and the current input X t splicing; C f is the weight matrix, z f is the bias term;

[0023] Memory cell state update: Final memory cell state C t is the previous memory state C t-1 and the candidate state K at the current moment t Update together and C t Input to output gate:

[0024]

[0025] Furthermore, the output gate is connected to the memory cell state C t Filter and adjust, and output the hidden state information h at the current time t t ;

[0026] h t =o t tanh(C t )

[0027] o t =σ(C o ·[h t-1 ,X t ]+z o )

[0028] Among them, C o is the weight matrix, Z o is the bias term.

[0029] Furthermore, the multi-head attention mechanism layer is specifically used to:

[0030] The hidden state of the second-layer LSTM output is linearly transformed to obtain the query Q, key K and value; the query, key and value are used to calculate the attention output Attention; the outputs of all attention heads are spliced ​​and the final multi-head attention output is obtained through linear transformation; the hidden state of the second-layer LSTM output is weightedly concatenated with the multi-head attention mechanism, and the splicing result is passed to the fully connected layer for feature extraction.

[0031] Furthermore, the second layer LSTM output H (2) Perform a linear transformation to compute the query Q, key K, and value V:

[0032]

[0033] Among them, W Q , W K , W V is the weight matrix, and the attention output Attention is calculated using the query, key, and value; for each attention head i:

[0034]

[0035] Among them, the softmax function is:

[0036]

[0037] Among them, z i is the ith input value, is the input value z i The index of is the exponential sum of all input values;

[0038] The outputs of all heads are concatenated and linearly transformed to obtain the final multi-head attention output:

[0039] MultiHead(Q,K,V)=W O Concat(head1,head2,...,head h )

[0040] Among them, W O is the output weight matrix, Concat is concatenation; use weighted concatenation of the second layer LSTM output and multi-head attention mechanism:

[0041] H combined =H (2) +MultiHead(Q,K,V)

[0042] The combined output is passed into the fully connected layer for feature extraction.

[0043] Furthermore, the fully connected layer uses the following formula to transform H combined Perform feature extraction:

[0044] g t =W f ·H combined +a f

[0045] Among them, Wf and a f is the weight matrix and bias vector of the fully connected layer, g t is the extracted feature.

[0046] Furthermore, the loss function for training the long short-term memory neural network model is defined, and the loss function L is:

[0047]

[0048]

[0049] Among them, λ is the regularization hyperparameter, w i is the sample weight, σ is the standard deviation of the prediction error, is the predicted temperature of the ith cable core, The actual temperature of the i-th cable core, n represents the total number of cable cores.

[0050] Furthermore, the upper and lower temperature change thresholds ε during normal operation are established through the mean and standard deviation of the temperature change values ​​during normal operation of the cable core:

[0051] ε=Δμ s +2Δσ s

[0052] In the formula, Δμ s It represents the overall average value of the temperature change at every two adjacent moments when the cable core is working normally, σ s is the overall standard deviation of the temperature change value of the cable core at every two adjacent moments; then, ε is superimposed with the cable core temperature prediction value of the inversion model at time t-1 to obtain the dynamic threshold value of the cable core temperature at time t δ t :

[0053]

[0054]

[0055] In the formula, is the predicted value of the cable core temperature at time t-1, obtained by the cable core temperature inversion model (LSTM model based on multi-head attention mechanism); if the cable core temperature at the current moment Exceeding the dynamic warning threshold δ t , it is considered that the cable status may be abnormal.

[0056] Due to the adoption of the above technical solution, the present invention has the following advantages:

[0057] The present invention utilizes the cable surface temperature, current carrying capacity and material thermal conductivity, and takes into account the loss of ground loop current. The cable core temperature is calculated by constructing a long short-term memory neural network model. Compared with the commonly used temperature monitoring and threshold methods, it can combine multi-source data and infer the cable core temperature. It can more accurately judge the status of the cable and provide more scientific and effective guidance for cable maintenance and repair. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0059] Figure 1 A schematic flow chart of a cable core temperature dynamic early warning method based on multi-source parameter inversion according to an embodiment of the present invention (including a dual long short-term memory neural network model);

[0060] Figure 2 This is a schematic diagram of the cable core temperature inversion result of normal training according to an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of predicting abnormal cable core temperature according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention is further described in conjunction with the accompanying drawings and embodiments, and the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present invention.

[0063] In the prior art, traditional cable temperature prediction models usually use simple regression models, neural networks or shallow time series models for prediction, and the loss function is usually RMSE or MSE. These models have limitations when processing complex data with strong time dependence, and cannot effectively capture long-term dynamic trends and short-term rapid fluctuations. The two-layer LSTM (Long Short-Term Memory Network) is designed specifically for processing time series data and can capture both short-term and long-term dependencies of data. Through the two-layer structure, the model can more accurately understand the dynamic temperature changes of the cable: the first layer of LSTM is used to capture the temperature change characteristics in a short period of time (such as current carrying capacity fluctuations in a short period of time). The second layer of LSTM is used to capture the temperature change trend over a longer time range (such as the impact of thermal conductivity on long-term changes in ambient temperature). And the loss function that introduces weight adjustment and regularization terms is used for training, which can effectively reduce the sensitivity to outliers during calculation and improve the generalization ability of the model. This deep time series modeling capability and training loss function enable the solution to better handle complex temperature time series relationships compared to the prior art and provide more accurate temperature prediction.

[0064] In the existing technology, the model often only considers the surface temperature and current carrying capacity of the cable, while ignoring other factors that may affect the core temperature. The input parameters of the proposed method include cable surface temperature, current carrying capacity, material thermal conductivity and ground loop loss. Through the combination of multi-source data, the multi-variable input layer (including surface temperature, current carrying capacity, material thermal conductivity and ground loop loss) can help the model more comprehensively describe the working state of the cable, especially considering the effects of thermal conductivity and energy loss, which are key parameters that determine the core temperature of the cable. Through the collaborative analysis of these multi-dimensional input data, the model can learn the relationship and interaction between different variables and capture more complex influence patterns.

[0065] The cable monitoring system in the prior art generally uses a static temperature threshold, and issues an alarm when the cable core temperature exceeds a preset upper temperature limit. This static threshold does not take into account the complexity of actual working conditions and is prone to false alarms or missed alarms, especially when the cable operating conditions are complex and changeable.

[0066] The present invention introduces a cable core temperature threshold method based on a dynamic threshold:

[0067] The temperature change threshold during normal operation is established by the mean and standard deviation of the temperature change value of the cable core during normal operation, and is superimposed with the cable core temperature prediction value at the previous moment to form the dynamic warning threshold of the cable core temperature at the current moment. When the cable core temperature exceeds the dynamic threshold, the model triggers an action warning. This dynamic threshold can avoid the limitations brought by the static threshold setting and reduce the probability of false alarms or missed alarms.

[0068] See also Figure 1The present invention provides an embodiment of a cable core temperature dynamic early warning method based on multi-source parameter inversion, which comprises the following steps:

[0069] Step 1: Establish a long short-term memory neural network model consisting of seven layers of networks. The first layer is used to input the air convection coefficient α of the cable, the surface temperature T, and the average working temperature T W , average ambient temperature T E , current carrying capacity I, real-time skin coefficient ρ, thermal resistance R and ground loop current loss P:

[0070] X t ={X1,X2,...,X t}

[0071] Among them, X t =[α t ,T t ,T Wt ,T Et ,I t ,ρ t ,R t ,P t ].

[0072] The role of the LSTM layer is to capture long-term dependencies in time series. The core of LSTM lies in its unit structure, which includes input gate, forget gate and output gate. Let the hidden state of this layer be h t , the state update formula is as follows:

[0073] The function of the input gate is to control how much new information should be added to the memory cell at the current moment. The output of the input gate is also a value between 0 and 1, which is used to determine which new information is added to the memory cell.

[0074]

[0075] in, is the output of the input gate at time t, h t-1 is the hidden state at the previous time t-1, C i is the weight matrix, z i is the bias term, σ is the sigmoid function;

[0076] In addition, the input gate generates a candidate memory cell K t , which represents the new information content that can be introduced at the current moment to generate a new candidate state.

[0077] K t =tanh(C K ·[h t-1 ,X t ]+z K )

[0078] Among them, the tanh function is:

[0079]

[0080] Among them, tanh is the hyperbolic tangent function, C K is the weight matrix, z K is the bias term. e is a natural constant, approximately equal to 2.718, and x is the input value.

[0081] Forget gate: The function of the forget gate is to determine which information in the memory unit of the previous moment needs to be retained and which needs to be discarded. The output of the forget gate is a value between 0 and 1, where 0 means completely forgotten and 1 means completely retained.

[0082] f t =σ(C f ·[h t-1 ,X t ]+z f )

[0083] Among them, f t is the information that needs to be retained by the forget gate, σ is the sigmoid function; [h t-1 ,X t ] is the hidden state h at the previous time t-1 t-1 and the current input X t splicing; C f is the weight matrix, z f is the bias term;

[0084]

[0085] Memory cell state update: Final memory cell state C t is the previous memory state C t-1 and the candidate state K at the current moment t Update together and C t Input to output gate:

[0086]

[0087] Output gate: The function of the output gate is to determine the hidden state h at the current time t t The output of memory cell C t Filter and adjust, and output the hidden state information at the current moment.

[0088] o t =σ(C o ·[h t-1 ,X t ]+z o )

[0089] The hidden state at the current moment is calculated by the following formula:

[0090] h t =o t tanh(C t )

[0091] Among them, C o is the weight matrix, Z o is the bias term, tanh(C t ) is the output of the current memory unit after being activated by the tanh function.

[0092] The above can be summarized as follows:

[0093]

[0094] The ReLU activation function is then used to optimize the training:

[0095]

[0096] The formula for the ReLU activation function is as follows:

[0097]

[0098] When using ReLU, the neuron is activated only when the input is positive, which means it produces output only for positive values, thus creating sparsity.

[0099] The Dropout layer is used to reduce the effect of overfitting.

[0100]

[0101] The function of Dropout is:

[0102] For each neuron h, it is randomly discarded with probability p in each iteration, and the probability of retaining it is 1-p;

[0103]

[0104] During the testing phase, all neurons are retained and their outputs are scaled to ensure consistency in output during training and testing.

[0105] The neurons that have passed through Dropout are then put into the second layer of LSTM to further capture the complex patterns and features in the data. The structure of this layer is the same as the first layer.

[0106]

[0107] Hidden state of the second LSTM layer output Input into the multi-head attention mechanism and define the output of the second layer LSTM as H (2) :

[0108]

[0109] Where T is the time step, is the hidden state at the Tth time step.

[0110] The output of the second LSTM layer is linearly transformed to calculate the query Q, key K and value V:

[0111]

[0112] Among them, W Q , W K , W V is the weight matrix.

[0113] Compute the attention output Attention using the query, key, and value. For each attention head i:

[0114]

[0115] Among them, the softmax function is:

[0116]

[0117] Among them, z i is the ith input value, is the input value z i The index of is the exponential sum of all input values;

[0118] The outputs of all heads are concatenated and linearly transformed to obtain the final multi-head attention output:

[0119] MultiHead(Q,K,V)=W O Concat(head1,head2,...,head h )

[0120] Among them, W O is the output weight matrix, Concat is concatenation, that is, assuming that the shape of A is (m, n) and the shape of B is (m, p), the result of the Concat operation is a tensor with a shape of (m, n+p).

[0121] Then use the weighted concatenation of the second layer LSTM output and the multi-head attention mechanism:

[0122] H combined =H (2) +MultiHead(Q,K,V)

[0123] The combined output is passed to the fully connected layer for feature extraction:

[0124] g t =W f ·H combined +a f

[0125] Among them, W f and a f is the weight matrix and bias vector of the fully connected layer, g t is the extracted feature. Finally, the output feature g of the fully connected layer t is input to the regression layer to predict the cable core temperature T 芯，预测 :

[0126] T 芯，预测 =W r ·g t +a r

[0127] Among them, W and a r is the weight matrix and bias vector of the regression layer, and the final output is the predicted cable core temperature T 芯，预测 .

[0128] About long short-term memory neural network model training:

[0129] In order to train the long short-term memory neural network model, it is necessary to define a loss function. The loss function is:

[0130]

[0131]

[0132] Where λ is the regularization hyperparameter, w i is the sample weight, σ is the standard deviation of the prediction error, is the predicted temperature of the ith cable core, The actual temperature of the i-th cable core, n represents the total number of cable cores.

[0133] The loss function can effectively reduce the sensitivity to outliers during calculation, while improving the generalization ability of the model by introducing weight adjustment and regularization terms.

[0134] Step 2: Input the measured surface temperature T, current carrying capacity I, material thermal conductivity ρ, and ground loop loss P of the cable under actual working conditions (the cable can be placed in actual working conditions through a temperature increase experiment) into the long short-term memory neural network model to obtain the predicted cable core temperature T 芯，预测 ;

[0135] T 芯，预测=f(T,I,ρ,P)

[0136] Step 3: Through the mean and standard deviation of the temperature change value of the cable core during normal operation, the upper and lower temperature change thresholds ε during normal operation are established:

[0137] ε=Δμ s +2Δσ s

[0138] In the formula, Δμ s It represents the overall average value of the temperature change at every two adjacent moments when the cable core is working normally, σ s It is the overall standard deviation of the temperature change value of the cable core at every two adjacent moments. Superimpose ε and the cable core temperature prediction value of the inversion model at time t-1 to obtain the dynamic threshold value δ of the cable core temperature at time t. t :

[0139]

[0140]

[0141] In the formula, It is the predicted value of the cable core temperature at time t-1, which is obtained by the cable core temperature inversion model.

[0142] Finally, by calculating the predicted cable core temperature at the current time t The predicted upper and lower dynamic thresholds δ t Compare to determine whether the cable core is working properly:

[0143]

[0144] If the predicted temperature at the current time t In the interval [δ t,lower ,δ t,upper ], it is considered that the cable status may be abnormal.

[0145] At a certain cable line crossing, sensors are used to measure the surface temperature T, current carrying capacity I, material thermal conductivity ρ, and ground loop current loss P. The parameters are normalized and input into the trained long short-term memory neural network model to predict the cable core temperature and calculate the cable core temperature. Figure 2 It can be seen that the difference between the predicted temperature and the actual temperature of the cable core is small, the RMSE is within 0.1628℃ and there is no abnormal alarm, which means that the result is very accurate.

[0146] Then the LSTM neural network model is applied to the cable with a problem in the core, and its dynamic threshold is observed. Figure 3As shown, it can be seen that the abnormal temperature alarm occurred five times, indicating that there is a problem with the cable core of the cable and it should be repaired. In addition, the predicted temperature can also be compared with the maximum allowable temperature of the cable core to issue an early warning.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A cable core temperature dynamic early warning method based on multi-source parameter inversion, characterized in that: include: The surface temperature, air convection coefficient, thermal resistance, average ambient temperature, average operating temperature, current carrying capacity, ground loop current loss and skin coefficient of the cable under actual working conditions are input into a double-layer long short-term memory neural network model based on a multi-head attention mechanism to obtain the predicted cable core temperature at the current moment; The temperature change threshold during normal operation is established through the mean and standard deviation of the temperature change value of the cable core during normal operation, and is superimposed with the cable core temperature prediction value at the previous moment to form the dynamic warning threshold of the cable core temperature at the current moment; if the cable core temperature at the current moment exceeds the dynamic warning threshold, it is considered that the cable status may be abnormal.

2. The method according to claim 1, characterized in that The long short-term memory neural network model is composed of a sequence input layer, a first LSTM layer, an activation function layer, a Dropout layer, a second LSTM layer, a multi-head attention mechanism layer, a fully connected layer and a regression layer connected in sequence. A loss function is used during training, and weight adjustment and regularization terms are introduced to improve the generalization ability of the model. The sequence input layer is used to input the air convection coefficient α, surface temperature T, and average operating temperature T of the cable. W , average ambient temperature T E The time series X at the current time t is composed of the current carrying capacity I, the real-time skin coefficient ρ, the thermal resistance R and the ground loop loss P t ; The first LSTM layer includes a memory unit, which includes an input gate, a forget gate, and an output gate. The input gate is used to control the new information added to the memory unit at the current moment. The output of the input gate is used to determine the new information added to the memory unit. The input gate generates a candidate memory unit, which represents the new information content that can be introduced at the current moment, and generates a new candidate state. The forget gate is used to determine the information that needs to be retained and discarded in the memory unit at the previous moment. The output gate is used to determine the output of the hidden state at the current moment. The Relu activation function of the activation function layer is used to optimize the hidden state output by the first LSTM layer for training; the Dropout layer is used to reduce the overfitting effect of the hidden state after optimization training and output neurons; the second LSTM layer is used to receive the neurons output by the Dropout layer and output the hidden state; the multi-head attention mechanism layer is used to receive the hidden state output by the second LSTM layer, and use weighted splicing of the hidden state output by the second LSTM layer and the multi-head attention mechanism, and output the splicing result; the fully connected layer is used to receive the splicing result and output the predicted cable core temperature.

3. The method according to claim 2, characterized in that The input gate uses the following formula to calculate the received time series X t ={X1,X2,…,X t } for processing: in, is the output of the input gate at time t, h t-1 is the hidden state at the previous time t-1, C i is the weight matrix, z i is the bias term, σ is the sigmoid function; the input gate generates a candidate memory unit, which represents the new information content that can be introduced at the current moment, and generates a new candidate state K t ; K t =tanh(C K ·[h t-1 ,X t ]+z K ) Among them, tanh is the hyperbolic tangent function, C K is the weight matrix, z K is the bias term.

4. The method according to claim 3, characterized in that The forget gate uses the following formula to calculate the received time series X t ={X1,X2,…,X t } for processing: f t =σ(C f ·[h t-1 ,X t ]+z f ) Among them, f t is the information that needs to be retained by the forget gate, σ is the sigmoid function; [h t-1 ,X t ] is the hidden state h at the previous time t-1 t-1 and the current input X t splicing; C f is the weight matrix, z f is the bias term; Memory cell state update: Final memory cell state C t is the previous memory state C t-1 and the candidate state K at the current moment t Update together and C t Input to output gate:

5. The method according to claim 4, characterized in that The output gate passes through the memory cell state C t Filter and adjust, and output the hidden state information h at the current time t t ; h t =o t ·tanh(C t ) o t =σ(C o ·[h t-1 ,X t ]+z o ) Among them, C o is the weight matrix, Z o is the bias term.

6. The method according to claim 2, characterized in that The multi-head attention mechanism layer is specifically used for: The hidden state of the second-layer LSTM output is linearly transformed to obtain the query Q, key K and value; the query, key and value are used to calculate the attention output Attention; the outputs of all attention heads are concatenated and the final multi-head attention output is obtained through linear transformation; the hidden state of the second-layer LSTM output is weighted concatenated with the multi-head attention mechanism, and the concatenation result is passed to the fully connected layer for feature extraction.

7. The method according to claim 6, characterized in that The second layer LSTM output H (2) Perform a linear transformation to compute the query Q, key K, and value V: Among them, W Q , W K , W V is the weight matrix, and the attention output Attention is calculated using the query, key, and value; for each attention head i: Among them, the softmax function is: Among them, z i is the ith input value, is the input value z i The index of is the exponential sum of all input values; Concatenate the outputs of all heads and use linear transformation to get the final multi-head attention output MultiHead(Q,K,V)=W O ·Concat(head1,head2,...,head h ) Among them, W O is the output weight matrix, Concat is the concatenation where the calculation of each head is; Use weighted concatenation of the second layer LSTM output and multi-head attention mechanism: H combined =H (2) +MultiHead(Q,K,V) The combined output is passed into the fully connected layer for feature extraction.

8. The method according to claim 7, characterized in that The fully connected layer uses the following formula to transform H combined Perform feature extraction: g t =W f ·H combined +a f Among them, W f and a f is the weight matrix and bias vector of the fully connected layer, g t is the extracted feature.

9. The method according to claim 1, characterized in that: Define the loss function for training the long short-term memory neural network model. The loss function L is: Among them, λ is the regularization hyperparameter, w i is the sample weight, σ is the standard deviation of the prediction error, is the predicted temperature of the ith cable core, The actual temperature of the i-th cable core, n represents the total number of cable cores.

10. The method according to claim 1, characterized in that The upper and lower temperature change thresholds ε during normal operation are established through the mean and standard deviation of the temperature change value of the cable core during normal operation: e=Dm s +2Ds s In the formula, Δμ s It represents the overall average value of the temperature change at every two adjacent moments when the cable core is working normally, σ s is the overall standard deviation of the temperature change value of the cable core at every two adjacent moments; then, ε is superimposed with the cable core temperature prediction value of the inversion model at time t-1 to obtain the dynamic threshold value of the cable core temperature at time t δ t : In the formula, is the predicted value of the cable core temperature at time t-1, obtained by the cable core temperature inversion model; If the cable core temperature at the current moment Exceeding the dynamic warning threshold δ t , it is considered that the cable status may be abnormal.

Images