A deep learning-based intelligent control method for power distribution cabinets

Abstract

The present invention discloses an intelligent control method for a power distribution cabinet based on deep learning, and relates to the technical field of intelligent power distribution. The present invention constructs a deep recursive neural network model based on the output of renewable energy power generation, power grid operation status and environmental information collected by sensors, extracts the time series characteristics of power distribution parameters in combination with historical and real-time data, predicts short-term load demand and change trend, and introduces gradually refined error quantization and optimization algorithms to ensure rapid convergence of the model and high prediction accuracy; based on the load prediction results, a deep reinforcement learning algorithm is used to generate an optimal control strategy, and the power distribution efficiency and stability are maximized by defining the state space, action space and reward function, while dynamically adjusting the output power of each branch to ensure the consistency of load distribution; in the actual execution process, the total power of the output parameters is decomposed and verified, and the rationality of the distribution logic is further improved in combination with iterative adjustment, and the power factor is optimized to balance the reactive power when necessary.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent power distribution, and in particular to an intelligent control method of a power distribution cabinet based on deep learning. Background Art

[0002] Renewable energy generation is intermittent and random. For example, photovoltaic power generation will cause rapid fluctuations in output power due to changes in sunlight intensity or cloud cover, while wind power systems will change due to unstable wind speeds. With the access of distributed renewable energy sources such as wind and solar energy, the load volatility of the power system has significantly increased.

[0003] Traditional intelligent control methods for distribution cabinets are mainly used in scenarios with relatively stable loads such as industry, commerce or residences. Their design ideas focus on simple control under preset rules. Specifically, by monitoring parameters such as voltage and current and comparing them with thresholds, warnings are triggered or switching operations are performed when abnormalities are found. This method relies on centralized management and is regulated by a unified monitoring system. It can meet operating requirements under normal load conditions, but when faced with new energy sources, it is difficult to quickly adapt to their load conditions, and the ability to predict complex fluctuation patterns is insufficient. The control logic is also limited to simple fixed threshold rules, and the power supply parameters cannot be dynamically adjusted.

[0004] To this end, some traditional solutions increase the sampling frequency of sensors, expand the monitoring coverage, and optimize the control threshold, which to a certain extent improves the sensitivity and adaptability of control, but still cannot fundamentally solve the core problem of new energy load fluctuations. In complex scenarios where the load changes drastically or multiple sources of fluctuations overlap, the effect is limited. Therefore, a deep learning-based intelligent control method for distribution cabinets is urgently needed to solve such problems. Summary of the invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a deep learning-based intelligent control method for distribution cabinets to solve the problems that traditional control methods rely on centralized management, are difficult to adapt to the rapidly changing load conditions of new energy, have insufficient prediction capabilities for complex fluctuation patterns, and the control logic is limited to simple fixed threshold rules, making it impossible to dynamically adjust power supply parameters.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: The embodiment of the present invention provides a distribution cabinet intelligent control method based on deep learning, which includes: Step S1, placing sensors in the power distribution cabinet to collect power distribution parameters, including power generation parameters, power grid operation status and environmental information; Step S2, building a deep recursive neural LSTM model to analyze the power distribution parameters and predict future load demand and change trends; Step S3, generating a control strategy based on the load demand and change trend predicted in step S2; Step S4, adjusting the output parameters of the distribution cabinet shunt based on the control strategy and implementing the load distribution logic; Step S5, collect the power distribution parameters adjusted in step S4, quantify the deviation from the predicted target, combine the environmental data and operation errors, enhance the training set of the deep learning model, and re-optimize the model weight parameters.

[0008] As a preferred solution of the deep learning-based intelligent control method for power distribution cabinets described in the present invention, the power generation parameters include the power generation output of photovoltaic or wind power new energy, The grid operation status includes voltage, current and power factor. The environmental information includes the temperature, humidity and wind speed of the environment in which the power distribution cabinet is located.

[0009] As a preferred solution of the intelligent control method of a power distribution cabinet based on deep learning described in the present invention, the LSTM model is constructed as follows: The LSTM model is trained based on the historical distribution parameters of renewable energy power generation. The real-time collected distribution parameters are input into the trained LSTM model to analyze short-term load demand and explore possible change trends.

[0010] As a preferred solution of the intelligent control method of power distribution cabinet based on deep learning described in the present invention, the steps of constructing a deep recursive neural LSTM model, analyzing the power distribution parameters, and predicting future load demand and change trends are as follows: The historical power distribution parameters and the real-time collected power distribution parameters are normalized, and the calculation formula is: , in, Represents the original power distribution parameters, represents the normalized power distribution parameters, Minimum value of power distribution parameters, The maximum value of the power distribution parameter, Divide the normalized data into time windows and define the input sequence and target sequence : , , in, represents the input sequence within the time window, represents the target sequence, represents the current time step, represents the time window length of the input sequence, represents the prediction step length of the target sequence, Indicates that from the current time step Previous The power distribution parameters of the time step, Indicates that from the current time step Next The prediction target of time steps is ; At each time step, the input data is processed by the LSTM model. The internal calculations of the model include: Forget Gate: , Input Gate: , Candidate status: , Update cell status: , Output Gate: , Hidden layer state: , in, represents the activation value of the forget gate, represents the activation value of the input gate, represents the candidate state value, Represents the cell state value, represents the activation value of the output gate, represents the hidden layer state output, Indicates the input power distribution parameters for the current time step, represents the hidden layer state output at the previous time step, represents the cell state value at the previous time step, represents the Sigmoid activation function, represents the hyperbolic tangent activation function, represents the weight matrix, Indicates the bias value, Represents element-by-element multiplication operation; The last time step of the LSTM model The hidden state of Mapped to predicted value through fully connected layer , the mapping process is: , in, represents the predicted output sequence, represents the weight matrix of the fully connected layer, represents the bias term of the fully connected layer; The mean square error (MSE) is used as the loss function to measure the prediction accuracy of the model. The loss function is defined as: , in, represents the loss value, represents the number of samples, represents the prediction step length, Indicates samples from time step Start to go back The actual value of the time step, Indicates samples from time step Start to go back The predicted value of time steps; Calculate the gradient of the loss function to the parameters through back propagation, and use the Adam optimization algorithm to adjust the weights and bias values .

[0011] As a preferred solution of the deep learning-based intelligent control method for a power distribution cabinet described in the present invention, the control strategy includes: Based on the predicted future load demand and changing trend, the deep reinforcement learning DQN algorithm is used to calculate the voltage and current output of each distribution cabinet branch; and the power factor optimization algorithm is executed in the fluctuating load.

[0012] As a preferred solution of the intelligent control method of a power distribution cabinet based on deep learning described in the present invention, the step of generating a control strategy based on the load demand and change trend predicted in step S2 is: The future load demand and change trend predicted in step S2 are classified, and the load states are defined as high load, medium load and low load. The load states are classified using the threshold method. The classification formula is: like ,but , like ,but , like ,but , in, Represents the time step Load status classification, represents the time step predicted by step S2 The load demand, and are the classification thresholds for high load and low load, respectively; The deep reinforcement learning algorithm DQN is used to generate the voltage and current output control strategy for each distribution cabinet branch. The optimization steps of the control strategy include: Defining the state space : , in, Represents the time step The state space of Indicates the past The load forecast value of the time step, Represents the time step The environmental information vector includes temperature, humidity and wind speed. Defining the action space : , in, Represents the time step The action vector of and Represents the time step No. The voltage and current control values ​​of each branch, Indicates the total number of branches in the distribution cabinet. Defining the reward function : , in, Represents the time step The reward value, Indicates The actual load output power of each branch is Represents the time step The load forecast value, is the weight coefficient, Indicates the current power factor, represents the target power factor, The Q value update formula is: , in, Indicates status and actions The Q value, represents the learning rate, represents the discount factor, Represents the reward value, Indicates the next state The maximum Q value of After multiple rounds of training, the optimal Q value strategy is used to generate time steps The optimal control strategy is generated as follows: , in, Represents the time step The optimal control action is It means choosing the action with the largest Q value among all possible actions.

[0013] As a preferred solution of the deep learning-based intelligent control method for a power distribution cabinet described in the present invention, the steps of adjusting the output parameters of the power distribution cabinet branch based on the control strategy and implementing the load distribution logic are as follows: Using the optimal control action generated in step S3 , decompose it into the voltage and current output of each branch, the decomposition formula is: , in, Represents the time step The optimal control strategy for Represents the time step No. The optimal voltage output value of each branch, Represents the time step No. The optimal current output value of each branch, Indicates the total number of branches in the distribution cabinet. The output power of each branch is adjusted according to the control strategy, and the branch load distribution formula is defined as: , in, Represents the time step No. The output power of each branch, Calculate the total split output power , the calculation formula is: , And verify that the following conditions are met: , in, Represents the total output power of all branches, Represents the time step The load forecast value, Indicates the tolerance error range, If the conditions are not met, the voltage and current outputs are reallocated, and the adjustment formula for the reallocation is: , , in, and Represent the adjusted voltage and current respectively, and To adjust the factor, it is optimized through iterative calculation.

[0014] As a preferred solution of the deep learning-based intelligent control method for a power distribution cabinet described in the present invention, the step of implementing the load distribution logic includes: The adjusted shunt voltage and current values ​​are sent down and shunt output adjustment is performed. The actual output parameters are: , and record the actual output power at the same time; Further optimize the power factor under fluctuating load conditions , the optimization formula is: , in, Represents the time step The target power factor, Represents the time step of reactive power.

[0015] As a preferred solution of the intelligent control method of a power distribution cabinet based on deep learning described in the present invention, the step of quantifying the deviation from the predicted target, combining environmental data and operation errors, and enhancing the training set of the deep learning model is as follows: Record the actual output parameters of each branch after step S4 is implemented, including the actual voltage , Current and output power , calculate the total branch output power , the calculation formula is: , in, Represents the time step The actual total output power of all branches, Represents the time step No. The actual output power of each branch is Indicates the total number of branches in the distribution cabinet; Calculate the deviation between the load prediction value and the actual total output power. The calculation formula is: , in, Represents the time step The load deviation value, Represents the time step The actual total output power, Represents the time step The load forecast value, like , If it is the tolerance error range, the error is recorded for subsequent model optimization; Collect time steps Environmental data and running error , define the data augmentation input vector: , in, Represents the time step The data augmentation input vector is Represents environmental data, including temperature, humidity, and wind speed, Indicates operating errors, including power factor deviation and power imbalance rate between branches; The new data vector Add to training set : , in, represents the training set of the deep learning model, Represents the time step The data vector.

[0016] As a preferred solution of the deep learning-based intelligent control method for a power distribution cabinet described in the present invention, the step of re-optimizing the model weight parameters is as follows: Retrain the deep learning model using the expanded training set and update the weight parameters and bias parameters , the optimization process includes: Define the loss function: , Compute gradients and update parameters: , in, represents the loss value, represents the number of training samples, represents the prediction step length, Indicates time step in samples The actual value of represents the predicted value, represents all model parameters, represents the learning rate; Using a validation set Evaluate the performance of the updated model and calculate the mean absolute error, which is calculated as: , in, represents the mean absolute error, represents the number of samples in the validation set, Indicates the actual value, Represents the predicted value, if If the value exceeds the set threshold, the hyperparameters are further adjusted.

[0017] The beneficial effects of the present invention are: The present invention constructs a deep recursive neural network model based on the output of renewable energy power generation, power grid operation status and environmental information collected by sensors, extracts the time series characteristics of distribution parameters in combination with historical and real-time data, predicts short-term load demand and change trends, and still has good dynamic adaptability in complex load fluctuation scenarios. At the same time, a gradually refined error quantization and optimization algorithm are introduced to ensure rapid convergence of the model and high prediction accuracy. Based on the load prediction results, a deep reinforcement learning algorithm is used to generate an optimal control strategy, which maximizes the distribution efficiency and stability by defining the state space, action space and reward function, and dynamically adjusts the output power of each branch to ensure the consistency of load distribution. In the actual execution process, the strategy decomposes and verifies the total power of the output parameters, further improves the rationality of the distribution logic by combining iterative adjustment, and optimizes the power factor to balance reactive power when necessary. Finally, actual operation data is collected, the prediction deviation is quantified, and the training set is expanded in combination with environmental information. The model parameters are optimized by retraining to improve the adaptability and generalization ability of the model under dynamic loads and extreme conditions, thereby achieving higher energy utilization efficiency and power supply stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0019] Figure 1 This is a flow chart of the intelligent control method of power distribution cabinet based on deep learning of the present invention. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0021] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, reference Figure 1 , this embodiment provides a distribution cabinet intelligent control method based on deep learning, comprising the following steps: Step S1, placing sensors in the power distribution cabinet to collect power distribution parameters, including power generation parameters, power grid operation status and environmental information; The power generation parameters include the power generation output of photovoltaic or wind power new energy, The grid operation status includes voltage, current and power factor. The environmental information includes the temperature, humidity and wind speed of the environment in which the power distribution cabinet is located; Step S2, building a deep recursive neural LSTM model to analyze the power distribution parameters and predict future load demand and change trends; The LSTM model is constructed as follows: The LSTM model is trained based on the historical distribution parameters of renewable energy generation. The real-time distribution parameters are input into the trained LSTM model to analyze short-term load demand and explore possible change trends. The steps of constructing a deep recursive neural LSTM model, analyzing the power distribution parameters, and predicting future load demand and change trends are as follows: The historical power distribution parameters and the real-time collected power distribution parameters are normalized, and the calculation formula is: , in, Represents the original power distribution parameters, represents the normalized power distribution parameters, Minimum value of power distribution parameters, The maximum value of the power distribution parameter, Divide the normalized data into time windows and define the input sequence and target sequence : , , in, represents the input sequence within the time window, represents the target sequence, represents the current time step, represents the time window length of the input sequence, represents the prediction step length of the target sequence, Indicates that from the current time step Previous The power distribution parameters of the time step, Indicates that from the current time step Next The prediction target of time steps is ; At each time step, the input data is processed by the LSTM model. The internal calculations of the model include: Forget Gate: , Input Gate: , Candidate status: , Update cell status: , Output Gate: , Hidden layer state: , in, represents the activation value of the forget gate, represents the activation value of the input gate, represents the candidate state value, Represents the cell state value, represents the activation value of the output gate, represents the hidden layer state output, Indicates the input power distribution parameters for the current time step, represents the hidden layer state output at the previous time step, represents the cell state value at the previous time step, represents the Sigmoid activation function, represents the hyperbolic tangent activation function, represents the weight matrix, Indicates the bias value, Represents element-by-element multiplication operation; The last time step of the LSTM model The hidden state of Mapped to predicted value through fully connected layer , the mapping process is: , in, represents the predicted output sequence, represents the weight matrix of the fully connected layer, represents the bias term of the fully connected layer; The mean square error (MSE) is used as the loss function to measure the prediction accuracy of the model. The loss function is defined as: , in, represents the loss value, represents the number of samples, represents the prediction step length, Indicates samples from time step Start to go back The actual value of the time step, Indicates samples from time step Start to go back The predicted value of time steps; Calculate the gradient of the loss function to the parameters through back propagation, and use the Adam optimization algorithm to adjust the weights and bias values ;

[0024] Specifically, in this step, the timing characteristics of the distribution parameters are extracted through normalization, time series segmentation and cyclic calculation of the LSTM model. The LSTM model coordinates the flow of information through the forget gate, input gate and output gate to predict future load demand and change trends. The loss function introduces a step-by-step prediction refinement error quantization method to further improve the accuracy of model training, and combines it with the optimization algorithm to ensure rapid convergence of the model.

[0025] Step S3, generating a control strategy based on the load demand and change trend predicted in step S2; The control strategy includes: Based on the predicted future load demand and change trend, the deep reinforcement learning DQN algorithm is used to calculate the voltage and current output of each distribution cabinet branch; and the power factor optimization algorithm is executed in the fluctuating load; The step of generating a control strategy based on the load demand and change trend predicted in step S2 is: The future load demand and change trend predicted in step S2 are classified, and the load states are defined as high load, medium load and low load. The load states are classified using the threshold method. The classification formula is: like ,but , like ,but , like ,but , in, Represents the time step Load status classification, represents the time step predicted by step S2 The load demand, and are the classification thresholds for high load and low load, respectively; The deep reinforcement learning algorithm DQN is used to generate the voltage and current output control strategy for each distribution cabinet branch. The optimization steps of the control strategy include: Defining the state space : , in, Represents the time step The state space of Indicates the past The load forecast value of the time step, Represents the time step The environmental information vector includes temperature, humidity and wind speed. Defining the action space : , in, Represents the time step The action vector of and Represents the time step No. The voltage and current control values ​​of each branch, Indicates the total number of branches in the distribution cabinet. Defining the reward function : , in, Represents the time step The reward value, Indicates The actual load output power of each branch is Represents the time step The load forecast value, is the weight coefficient, Indicates the current power factor, represents the target power factor, The Q value update formula is: , in, Indicates status and actions The Q value, represents the learning rate, represents the discount factor, Represents the reward value, Indicates the next state The maximum Q value of After multiple rounds of training, the optimal Q value strategy is used to generate time steps The optimal control strategy is generated as follows: , in, Represents the time step The optimal control action is It means to select the action with the largest Q value among all possible actions;

[0026] Specifically, in this step, the load demand and change trend are classified and analyzed, and combined with deep reinforcement learning, a control strategy corresponding to the prediction result is generated. The load classification provides a reference for strategy optimization. The deep reinforcement learning algorithm optimizes the voltage and current output of the branch by maximizing the reward function, and provides a flexible response strategy for fluctuating loads in terms of power factor optimization, effectively improving the distribution efficiency.

[0027] Step S4, adjusting the output parameters of the distribution cabinet shunt based on the control strategy and implementing the load distribution logic; The steps of adjusting the output parameters of the distribution cabinet shunt based on the control strategy and implementing the load distribution logic are as follows: Using the optimal control action generated in step S3 , decompose it into the voltage and current output of each branch, the decomposition formula is: , in, Represents the time step The optimal control strategy for Represents the time step No. The optimal voltage output value of each branch, Represents the time step No. The optimal current output value of each branch, Indicates the total number of branches in the distribution cabinet. The output power of each branch is adjusted according to the control strategy, and the branch load distribution formula is defined as: , in, Represents the time step No. The output power of each branch, Calculate the total split output power , the calculation formula is: , And verify that the following conditions are met: , in, Represents the total output power of all branches, Represents the time step The load forecast value, Indicates the tolerance error range, If the conditions are not met, the voltage and current outputs are reallocated, and the adjustment formula for the reallocation is: , , in, and Represent the adjusted voltage and current respectively, and To adjust the factors, optimization is performed through iterative calculations; The steps of implementing the load distribution logic include: The adjusted shunt voltage and current values ​​are sent down and shunt output adjustment is performed. The actual output parameters are: , and record the actual output power at the same time; Further optimize the power factor under fluctuating load conditions , the optimization formula is: , in, Represents the time step The target power factor, Represents the time step of reactive power.

[0028] Specifically, deep reinforcement learning is used to generate control strategies and adjust the output parameters of each branch. At the same time, the consistency of the total load distribution is verified and the power factor optimization is implemented, which further improves the efficiency and stability of the distribution cabinet. Combined with dynamic adjustment and real-time feedback, it can effectively cope with complex load fluctuations.

[0029] Step S5, collecting the power distribution parameters adjusted in step S4, quantifying the deviation from the predicted target, combining environmental data with operation errors, enhancing the training set of the deep learning model, and re-optimizing the model weight parameters; The steps of quantifying the deviation from the predicted target, combining environmental data and operation errors, and enhancing the training set of the deep learning model are: Record the actual output parameters of each branch after step S4 is implemented, including the actual voltage , Current and output power , calculate the total branch output power , the calculation formula is: , in, Represents the time step The actual total output power of all branches, Represents the time step No. The actual output power of each branch is Indicates the total number of branches in the distribution cabinet; Calculate the deviation between the load prediction value and the actual total output power. The calculation formula is: , in, Represents the time step The load deviation value, Represents the time step The actual total output power, Represents the time step The load forecast value, like , If it is the tolerance error range, the error is recorded for subsequent model optimization; Collect time steps Environmental data and running error , define the data augmentation input vector: , in, Represents the time step The data augmentation input vector is Represents environmental data, including temperature, humidity, and wind speed, Indicates operating errors, including power factor deviation and power imbalance rate between branches; The new data vector Add to training set : , in, represents the training set of the deep learning model, Represents the time step The data vector.

[0030] The step of re-optimizing the model weight parameters is: Retrain the deep learning model using the expanded training set and update the weight parameters and bias parameters , the optimization process includes: Define the loss function: , Compute gradients and update parameters: , in, represents the loss value, represents the number of training samples, represents the prediction step length, Indicates time step in samples The actual value of represents the predicted value, represents all model parameters, represents the learning rate; Using a validation set Evaluate the performance of the updated model and calculate the mean absolute error, which is calculated as: , in, represents the mean absolute error, represents the number of samples in the validation set, Indicates the actual value, Represents the predicted value, if If the value exceeds the set threshold, the hyperparameters are further adjusted.

[0031] Specifically, based on actual operating data and environmental parameters, combined with load prediction deviations, an enhanced deep learning training set is constructed. By retraining the model and optimizing weight parameters, the model can better adapt to dynamic load changes and environmental influences.

[0032] 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 preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A distribution cabinet intelligent control method based on deep learning, characterized in that: include, Step S1, placing sensors in the power distribution cabinet to collect power distribution parameters, including power generation parameters, power grid operation status and environmental information; Step S2, building a deep recursive neural LSTM model to analyze the power distribution parameters and predict future load demand and change trends; Step S3, generating a control strategy based on the load demand and change trend predicted in step S2; Step S4, adjusting the output parameters of the distribution cabinet shunt based on the control strategy and implementing the load distribution logic; Step S5, collect the power distribution parameters adjusted in step S4, quantify the deviation from the predicted target, combine the environmental data and operation errors, enhance the training set of the deep learning model, and re-optimize the model weight parameters.

2. The method for intelligent control of a power distribution cabinet based on deep learning according to claim 1, characterized in that: The power generation parameters include the power generation output of photovoltaic or wind power new energy, The grid operation status includes voltage, current and power factor. The environmental information includes the temperature, humidity and wind speed of the environment in which the power distribution cabinet is located.

3. The method for intelligent control of a power distribution cabinet based on deep learning as claimed in claim 2, characterized in that: The LSTM model is constructed as follows: The LSTM model is trained based on the historical distribution parameters of renewable energy power generation. The real-time collected distribution parameters are input into the trained LSTM model to analyze short-term load demand and explore possible change trends.

4. The method for intelligent control of a power distribution cabinet based on deep learning as claimed in claim 3, characterized in that: The steps of constructing a deep recursive neural LSTM model, analyzing the power distribution parameters, and predicting future load demand and change trends are as follows: The historical power distribution parameters and the real-time collected power distribution parameters are normalized, and the calculation formula is: ， in, Represents the original power distribution parameters, represents the normalized power distribution parameters, Minimum value of power distribution parameters, The maximum value of the power distribution parameter, Divide the normalized data into time windows and define the input sequence and target sequence : ， ， in, represents the input sequence within the time window, represents the target sequence, represents the current time step, represents the time window length of the input sequence, represents the prediction step length of the target sequence, Indicates that from the current time step Previous The power distribution parameters of the time step, Indicates that from the current time step Next The prediction target of time steps is ; At each time step, the input data is processed by the LSTM model. The internal calculations of the model include: Forget Gate: ， Input Gate: ， Candidate status: ， Update cell status: ， Output Gate: ， Hidden layer state: ， in, represents the activation value of the forget gate, represents the activation value of the input gate, represents the candidate state value, Represents the cell state value, represents the activation value of the output gate, represents the hidden layer state output, Indicates the input power distribution parameters for the current time step, represents the hidden layer state output at the previous time step, represents the cell state value at the previous time step, represents the Sigmoid activation function, represents the hyperbolic tangent activation function, represents the weight matrix, Indicates the bias value, Represents element-by-element multiplication operation; The last time step of the LSTM model The hidden state Mapped to predicted value through fully connected layer , the mapping process is: ， in, represents the predicted output sequence, represents the weight matrix of the fully connected layer, represents the bias term of the fully connected layer; The mean square error (MSE) is used as the loss function to measure the prediction accuracy of the model. The loss function is defined as: ， in, represents the loss value, represents the number of samples, represents the prediction step length, Indicates samples from time step Start to go back The actual value of the time step, Indicates samples from time step Start to go back The predicted value of time steps; Calculate the gradient of the loss function to the parameters through back propagation, and use the Adam optimization algorithm to adjust the weights and bias values .

5. The method for intelligent control of a power distribution cabinet based on deep learning according to claim 4, characterized in that: The control strategy includes: Based on the predicted future load demand and changing trend, the deep reinforcement learning DQN algorithm is used to calculate the voltage and current output of each distribution cabinet branch; and the power factor optimization algorithm is executed in the fluctuating load.

6. The method for intelligent control of a power distribution cabinet based on deep learning according to claim 5, characterized in that: The step of generating a control strategy based on the load demand and change trend predicted in step S2 is: The future load demand and change trend predicted in step S2 are classified, and the load states are defined as high load, medium load and low load. The load states are classified using the threshold method. The classification formula is: like ,but , like ,but , like ,but , in, Represents the time step Load status classification, represents the time step predicted by step S2 The load demand, and are the classification thresholds for high load and low load, respectively; The deep reinforcement learning algorithm DQN is used to generate the voltage and current output control strategy for each distribution cabinet branch. The optimization steps of the control strategy include: Defining the state space : ， in, Represents the time step The state space of Indicates the past The load forecast value of the time step, Represents the time step The environmental information vector includes temperature, humidity and wind speed. Defining the action space : ， in, Represents the time step The action vector of and Represents the time step No. The voltage and current control values ​​of each branch, Indicates the total number of distribution cabinet branches. Defining the reward function : ， in, Represents the time step The reward value, Indicates The actual load output power of each branch is Represents the time step The load forecast value, is the weight coefficient, Indicates the current power factor, represents the target power factor, The Q value update formula is: ， in, Indicates status and actions The Q value, represents the learning rate, represents the discount factor, Represents the reward value, Indicates the next state The maximum Q value of After multiple rounds of training, the optimal Q value strategy is used to generate time steps The optimal control strategy is generated by: ， in, Represents the time step The optimal control action is It means choosing the action with the largest Q value among all possible actions.

7. The method for intelligent control of a power distribution cabinet based on deep learning according to claim 6, characterized in that: The steps of adjusting the output parameters of the distribution cabinet shunt based on the control strategy and implementing the load distribution logic are as follows: Using the optimal control action generated in step S3 , decompose it into the voltage and current output of each branch, the decomposition formula is: ， in, Represents the time step The optimal control strategy for Represents the time step No. The optimal voltage output value of each branch, Represents the time step No. The optimal current output value of each branch, Indicates the total number of distribution cabinet branches. The output power of each branch is adjusted according to the control strategy, and the branch load distribution formula is defined as: ， in, Represents the time step No. The output power of each branch, Calculate the total split output power , the calculation formula is: ， And verify that the following conditions are met: ， in, Represents the total output power of all branches, Represents the time step The load forecast value, Indicates the tolerance error range, If the conditions are not met, the voltage and current outputs are reallocated, and the adjustment formula for the reallocation is: ， ， in, and Represent the adjusted voltage and current respectively, and To adjust the factor, it is optimized through iterative calculation.

8. The method for intelligent control of a power distribution cabinet based on deep learning according to claim 7, characterized in that: The steps of implementing the load distribution logic include: The adjusted shunt voltage and current values ​​are sent down and shunt output adjustment is performed. The actual output parameters are: , and record the actual output power at the same time; Further optimize the power factor under fluctuating load conditions , the optimization formula is: ， in, Represents the time step The target power factor, Represents the time step of reactive power.

9. The method for intelligent control of a power distribution cabinet based on deep learning according to claim 8, characterized in that: The steps of quantifying the deviation from the predicted target, combining environmental data and operation errors, and enhancing the training set of the deep learning model are: Record the actual output parameters of each branch after step S4 is implemented, including the actual voltage , Current and output power , calculate the total branch output power , the calculation formula is: ， in, Represents the time step The actual total output power of all branches, Represents the time step No. The actual output power of each branch is Indicates the total number of branches in the distribution cabinet; Calculate the deviation between the load prediction value and the actual total output power. The calculation formula is: ， in, Represents the time step The load deviation value, Represents the time step The actual total output power, Represents the time step The load forecast value, like , If it is the tolerance error range, the error is recorded for subsequent model optimization; Collect time steps Environmental data and running error , define the data augmentation input vector: ， in, Represents the time step The data augmentation input vector is Represents environmental data, including temperature, humidity, and wind speed, Indicates operating errors, including power factor deviation and power imbalance rate between branches; The new data vector Add to training set : ， in, represents the training set of the deep learning model, Represents the time step The data vector.

10. The method for intelligent control of a power distribution cabinet based on deep learning according to claim 9, characterized in that: The step of re-optimizing the model weight parameters is: Retrain the deep learning model using the expanded training set and update the weight parameters and bias parameters , the optimization process includes: Define the loss function: ， Compute gradients and update parameters: ， in, represents the loss value, represents the number of training samples, represents the prediction step length, Indicates time step in samples The actual value of represents the predicted value, represents all model parameters, represents the learning rate; Using a validation set Evaluate the performance of the updated model and calculate the mean absolute error, which is calculated as: ， in, represents the mean absolute error, represents the number of samples in the validation set, Indicates the actual value, Represents the predicted value, if If the value exceeds the set threshold, the hyperparameters are further adjusted.

Images