Method and system for automatically controlling on-load voltage regulation of transformer

Abstract

The invention relates to the technical field of transformer voltage regulation, solves the technical problems that in the prior art, remarkable limitation exists when complex power system data is processed, and a nonlinear change rule influenced by multi-factor coupling is difficult to capture, and particularly relates to a method and system for automatically controlling on-load voltage regulation of a transformer. The method comprises the following steps: S1, generating a transformer database based on historical data of a transformer, and preprocessing original data in the transformer database to obtain standardized data. The time sequence decomposition algorithm splits the load into trend, period and random components, the adaptability to a complex load mode is enhanced, the prediction of the load prediction model can be more accurate through multi-layer multi-mode processing prediction in cooperation with hyper-parameter optimization, and the prediction accuracy and the processing efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of transformer voltage regulation technology, and in particular to a method and system for automatically controlling on-load voltage regulation of transformers. Background Technology

[0002] Automatic on-load tap changer (ATC) is a technology that dynamically adjusts the output voltage of a transformer while it is under load. It uses a built-in on-load tap changer to switch the tap positions of the transformer windings without disconnecting the circuit, thereby changing the turns ratio to regulate the voltage. The system is equipped with an automatic control device that can monitor voltage fluctuations at the grid or load end in real time. When the voltage deviates from the set value, it automatically triggers the tap changer to accurately compensate for the voltage deviation. While existing technology overcomes the shortcomings of traditional off-load tap changer, which requires power outages, and enables rapid voltage regulation under load fluctuations or grid faults, it still faces the following technical bottlenecks when dealing with complex power systems: insufficient ability to capture nonlinear voltage evolution trends; and a lack of effective multi-objective coordination mechanisms during voltage safety assessments, resulting in low control accuracy and decision-making efficiency. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for automatically controlling on-load voltage regulation of transformers. This solves the problems of insufficient ability to capture the nonlinear evolution trend of voltage and lack of an effective multi-objective coordination mechanism in existing technologies, which leads to low control accuracy and decision-making efficiency.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for automatically controlling on-load tap changing of a transformer, the method comprising the following steps: S1. Generate a transformer database based on the raw data of the transformer, and preprocess the raw data in the transformer database to obtain standardized data; S2. Based on standardized data, a load forecasting model is constructed using long short-term memory networks, attention mechanisms, and seasonal time series decomposition. The load forecasting model is then optimized to obtain an optimized forecasting model. Standardized data is then input into the optimized forecasting model to obtain multi-period load forecasting sequences and confidence intervals. S3. Based on the multi-period load forecast sequence, voltage safety assessment and key node identification are obtained sequentially through the voltage deviation amplitude coefficient formula, sensitivity coefficient matrix method and fuzzy comprehensive evaluation method. S4. Based on the use of transformers, safety specifications are generated. Within the scope of the safety specifications, the optimal voltage regulation strategy and strategy confidence are obtained by sequentially passing through state coding network, Bayesian deep Q network, multi-objective Bellman optimal equation and multi-criteria decision and confidence within the scope of the safety specifications. S5. Based on the optimal voltage regulation strategy and strategy confidence, specific voltage regulation commands are generated through command mapping. A state prediction model is constructed according to the specific voltage regulation commands. The voltage qualification rate and voltage regulation effect are evaluated by using the voltage qualification rate formula on the predicted voltage regulation state output by the state prediction model. S6. Based on the voltage qualification rate and voltage regulation effect, synchronization status information is generated by threshold comparison method and position difference comparison method. Based on the synchronization status information, multi-condition logic judgment is used to obtain the final control command. S7. Obtain the transformer's execution result based on the final control command, obtain updated historical data based on the execution result, perform deep optimization of the state prediction model based on the updated historical data and the current hyperparameters of the state prediction model to obtain a deep state model, and obtain updated control commands through the deep state model.

[0005] Furthermore, in step S1, the specific implementation steps are as follows: S11. Based on the raw data in the transformer database, outlier removal is performed using the Raida criterion to obtain qualified data; S12. Based on the qualified data, perform data smoothing and denoising using the wavelet threshold denoising method to obtain denoised data; S13. Obtain standardized data by Z-score standardization based on the denoised data.

[0006] Furthermore, in step S2, the specific implementation steps are as follows: S21. Construct a sliding window with a width of D and a sliding step size of m. Calculate the multi-scale temporal feature matrix of the standardized data within the sliding window using the time series feature extraction formula, and generate the temporal feature layer. S22. Generate a long short-term hidden state sequence using LSTM encoding based on the multi-scale temporal feature matrix, and generate an encoding layer; S23. Obtain the weight distribution through the attention mechanism based on the long and short-term hidden state sequences, obtain the context vector through weighted fusion of the weights based on the weight distribution, and construct the attention layer; S24. Based on the context vector, the vector is decomposed through seasonal time series decomposition to obtain the trend component, periodic component and residual component, and a vector decomposition layer is constructed. S25. Based on the trend component, periodic component, residual component and context vector, the initial multi-period load prediction sequence and initial confidence interval are obtained by sequentially processing the LSTM decoder and Monte Carlo Dropout. The prediction layer is then constructed, and the time series feature layer, encoding layer, attention layer, vector decomposition layer and prediction layer are combined into a load prediction model. S26. Based on the load forecasting model, the hyperparameters are optimized using the cross-validation method to obtain the optimized forecasting model. Standardized data is then input into the optimized forecasting model to obtain multi-period load forecasting sequences and confidence intervals.

[0007] Furthermore, in step S3, the specific implementation steps are as follows: S31. Calculate the voltage deviation amplitude coefficient using the voltage deviation amplitude coefficient formula based on the transformer's rated voltage and the multi-period load prediction sequence; S32. Obtain the power grid parameters from the transformer's equipment nameplate, and obtain the node voltage sensitivity matrix using the sensitivity coefficient matrix method based on the voltage deviation amplitude coefficient and the power grid parameters. S33. Based on the voltage deviation amplitude coefficient and the node voltage sensitivity matrix, the voltage safety assessment and key node identification are obtained through the fuzzy comprehensive evaluation method.

[0008] Furthermore, in step S4, the specific implementation steps are as follows: S41. Based on voltage safety assessment and key node identification, a low-dimensional state representation vector is calculated through a state coding network. S42. Generate the action interval based on the confidence interval and the safety specifications of the transformer. Obtain the posterior distribution of the Q value through a Bayesian deep Q network based on the low-dimensional state representation vector and the action interval. S43. Obtain the current state data of the transformer, and calculate the reward vector based on the state data and the posterior distribution of the Q value using a multi-objective reward function. S44. The Pareto optimal policy set is calculated using the multi-objective Bellman optimal equation based on the reward vector and the posterior distribution of the Q-value. S45. Based on the Pareto optimal strategy set, the optimal voltage regulation strategy and strategy confidence are obtained through multi-criteria decision-making and confidence calculation.

[0009] Furthermore, in step S5, the specific implementation steps are as follows: S51. The optimal voltage regulation strategy and strategy confidence are generated by the instruction mapping method to produce specific voltage regulation instructions. S52. Obtain the current state of the transformer. Based on the specific voltage regulation command and the current state, construct a state prediction model through a deep neural network. Input the specific voltage regulation command and the current state into the state prediction model to obtain the predicted voltage regulation state. S53. Obtain the voltage qualification rate using the voltage qualification rate formula based on the predicted voltage regulation status, and generate a voltage regulation effect evaluation based on the voltage qualification rate. If the voltage qualification rate is less than or equal to 95%, the voltage regulation effect is evaluated as unqualified. If the voltage qualification rate is greater than 95%, the voltage regulation effect is evaluated as qualified.

[0010] Furthermore, in step S6, the specific implementation steps are as follows: S61. Based on the voltage qualification rate and voltage regulation effect evaluation, the overcurrent blocking result is generated by the threshold comparison method. S62. Obtain the position data of multiple taps of the transformer, and synchronize the overcurrent blocking results according to the position data by using the position difference comparison method to obtain the synchronization status information. S63. The final control command is obtained by using multi-condition logic judgment based on the overcurrent blocking result, synchronization status information and predicted voltage regulation status.

[0011] Furthermore, in step S7, the specific implementation steps are as follows: S71. Complete the adjustment of the transformer according to the final control command and obtain the execution result after adjustment. Generate experience playback data through priority experience playback based on the execution result, and store the experience playback data as updated historical data in the transformer database. S72. Obtain the current hyperparameters of the state prediction model, and obtain a dynamic learning rate that can be dynamically adjusted by the historical gradient magnitude through the AdaGrad adaptive learning rate algorithm based on the updated historical data and the current hyperparameters. S73. Based on the dynamic learning rate, the state prediction model is deeply optimized through Bayesian deep Q-learning to obtain a deep state model. The deep state model replaces the state prediction model and the updated voltage regulation state is obtained. The updated control command is obtained through S6.

[0012] The technical solution also provides a system for the above-mentioned method of on-load tap changing of automatic control transformers, the system comprising: The standardization module is used to generate a transformer database based on historical transformer data, and to preprocess the raw data in the transformer database to obtain standardized data. The load forecasting module is used to construct a load forecasting model based on standardized data through long short-term memory network, attention mechanism and seasonal time series decomposition, optimize the load forecasting model to obtain an optimized forecasting model, and input the standardized data into the optimized forecasting model to obtain multi-period load forecasting sequences and confidence intervals. The evaluation module is used to obtain voltage safety assessment and key node identification based on the multi-period load forecast sequence through the voltage deviation amplitude coefficient formula, sensitivity coefficient matrix method and fuzzy comprehensive evaluation method. The strategy module is used to generate safety specifications based on the use of transformers. Within the scope of the safety specifications, based on voltage safety assessment and key node identification, the optimal voltage regulation strategy and strategy confidence are obtained sequentially through state coding network, Bayesian deep Q network, multi-objective Bellman optimal equation and multi-criteria decision and confidence. The effect module is used to generate specific voltage regulation commands through command mapping based on the optimal voltage regulation strategy and strategy confidence, construct a state prediction model according to the specific voltage regulation commands, and obtain the voltage qualification rate and voltage regulation effect evaluation from the predicted voltage regulation state output by the state prediction model through the voltage qualification rate formula. The control module is used to generate synchronization status information based on voltage qualification rate and voltage regulation effect through threshold comparison method and position difference comparison method, and to obtain the final control command based on the synchronization status information using multi-condition logic judgment. The periodic optimization module is used to obtain the transformer's execution result based on the final control command, obtain updated historical data based on the execution result, perform deep optimization of the state prediction model based on the updated historical data and the current hyperparameters of the state prediction model to obtain a deep state model, and obtain updated control commands through the deep state model.

[0013] By employing the above technical solution, the present invention provides a method and system for automatically controlling on-load tap changing of transformers, which has at least the following beneficial effects: 1. This invention captures the long-term time-dependent characteristics of load through an LSTM network and dynamically weights key time steps using an attention mechanism, significantly improving prediction accuracy by more than 25%. The time-series decomposition algorithm breaks down the load into trend, periodic, and random components, enabling the model to learn the changing patterns at different time scales and enhancing its adaptability to complex load patterns. Through multi-layer, multi-modal processing and hyperparameter optimization, the load prediction model can be made more accurate, improving both prediction accuracy and processing efficiency.

[0014] 2. This invention dynamically captures system state characteristics through a state coding network and combines it with a Bayesian deep Q-network to handle uncertainty, achieving dynamic optimization under the multi-objective Bellman optimal equation. It can balance multiple conflicting objectives such as stability and economy. The multi-criteria decision-making and confidence calculation mechanism further quantifies the advantages and reliability of the strategy, providing an intuitive basis for decision-making. This process has dynamic adaptability, uncertainty robustness, and multi-objective collaborative optimization capabilities, significantly improving the scientificity and safety of voltage regulation strategies and effectively reducing the risks and costs in transformer regulation caused by the single objective and simplified model of traditional methods.

[0015] 3. This invention uses a command mapping method to flexibly adjust the voltage regulation strategy according to the real-time needs of the system, avoiding the rigidity of fixed rules and the complex nonlinear relationship between the voltage regulation command and the system. This improves the accuracy of predicting the voltage regulation state, avoids the one-sidedness of a single threshold, and enhances the objectivity of the evaluation through a data-driven approach. The voltage regulation strategy is more in line with the actual working conditions, the prediction accuracy is significantly improved, and the stability and energy efficiency of voltage regulation are greatly improved.

[0016] 4. This invention, through a closed-loop learning paradigm of priority experience playback – adaptive learning rate – Bayesian deep optimization, can continuously learn from its own operational experience, dynamically adjust decision-making strategies and quantify decision risks, and continuously learn and optimize, significantly improving the system's control accuracy, robustness and long-term economy in real and complex environments. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of an automatic control method for on-load tap changing of a transformer according to the present invention; Figure 2 This is a structural block diagram of an automatic control transformer on-load tap changer system according to the present invention.

[0018] In the diagram: 1. Standardization module; 2. Load forecasting module; 3. Evaluation module; 4. Strategy module; 5. Effect module; 6. Control module; 7. Periodic optimization module. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0020] Due to the significant limitations of existing technologies in processing complex power system data, it is difficult to capture the nonlinear variation patterns of the coupled effects of multiple factors.

[0021] Example 1:

[0022] like Figure 1 As shown in the figure, this embodiment proposes a method for on-load tap change of an automatic control transformer, which can significantly improve prediction accuracy, enhance adaptability to complex load patterns, make the load prediction model more accurate, and improve prediction accuracy and processing efficiency. The method includes the following steps: S1. Generate a transformer database based on the raw transformer data, and preprocess the raw data in the transformer database to obtain standardized data. Historical data is prone to inaccuracy due to environmental factors affecting the data acquisition equipment during the acquisition process. To address these issues, the specific implementation steps are as follows: S11. Based on the raw data in the transformer database, outlier removal is performed using the Raida criterion to obtain qualified data. The Raida criterion is an outlier detection method based on statistical principles. It is mainly used to determine whether there are outlier data points in the dataset that significantly deviate from the normal range. Its core logic is to use the characteristics of normal distribution to determine a reasonable fluctuation range by calculating the mean and standard deviation of the data. Data points that exceed this range are considered outliers. The raw data includes relevant technical parameters of the transformer, such as transformer type, voltage, current, active power, reactive power, and load impedance angle. These technical parameters are time-stamped to facilitate calculations in subsequent steps.

[0023] S12. Based on the qualified data, the data is smoothed and denoised using the wavelet thresholding method to obtain denoised data. The wavelet thresholding method decomposes the signal into wavelets, applies threshold processing to the obtained wavelet coefficients (usually using a soft threshold function), and then performs wavelet reconstruction to obtain the denoised data. The wavelet thresholding method is a commonly used smoothing and denoising method, which will not be elaborated here.

[0024] S13. Standardized data is obtained by Z-score standardization based on the denoised data. Z-score standardization is a data preprocessing method that transforms the original data into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of differences in dimensions or distributions and making the data comparable. This invention uses wavelet denoising and Z-score standardization, which can effectively preserve data features while eliminating noise interference, improving the stability of subsequent model training.

[0025] S2. Based on standardized data, a load forecasting model is constructed using long short-term memory networks, attention mechanisms, and seasonal time-series decomposition. This model is then optimized to obtain an optimized forecasting model. Standardized data is input into the optimized model to obtain multi-period load forecast sequences and confidence intervals. Traditional load forecasting methods have significant limitations when processing complex power system data. They rely excessively on the linear characteristics or simple statistical patterns of historical data, making it difficult to capture the nonlinear changes in power load influenced by multiple factors such as weather, economic activity, and user behavior. To address these issues, the specific implementation steps are as follows: S21. Construct a sliding window with a width of D and a sliding step size of m. Calculate the multi-scale time series feature matrix of the standardized data within the sliding window using the time series feature extraction formula, and generate the time series feature layer. The time series feature extraction formula includes mean features, variance features, and trend features, which are obtained through the mean, variance, and trend calculation formulas, and will not be elaborated here.

[0026] S22. Generate a long short-term hidden state sequence based on the multi-scale temporal feature matrix using LSTM encoding, and generate an encoding layer. LSTM encoding is the Long Short-Term Memory network encoding. The calculation of the long short-term hidden state sequence of the temporal feature matrix is ​​performed through the input gate, forget gate, output gate, and cell state. This will not be elaborated here.

[0027] S23. Obtain the weight distribution based on the long and short-term hidden state sequences through the attention mechanism. Based on the weight distribution, obtain the context vector through weighted fusion of weights and construct the attention layer. The attention mechanism is a core method in deep learning, which aims to enable the model to dynamically focus on key information and ignore irrelevant content when processing sequential data or complex structures, thereby improving the model's performance and interpretability. By combining the attention mechanism with the LSTM model, the context capture capability can be enhanced, and the utilization rate of key information can be improved through dynamic weight allocation, thereby improving the interpretability of the model, adapting to multimodal and complex tasks, and improving processing efficiency.

[0028] S24. Based on the context vector, the vector is decomposed through seasonal time series decomposition to obtain the trend component, periodic component, and residual component, and a vector decomposition layer is constructed. Seasonal time series decomposition is a method that decomposes time series data into three core components: trend, seasonality, and residual. It aims to reveal long-term change patterns, periodic fluctuations, and random noise in the data. Its core objective is to simplify the analysis of complex time series data and provide a clearer perspective for tasks such as prediction and anomaly detection. It will not be elaborated on here.

[0029] S25. Based on the trend component, periodic component, residual component, and context vector, the initial multi-period load prediction sequence and initial confidence interval are obtained by sequentially processing the LSTM decoder and Monte Carlo Dropout. A prediction layer is then constructed, combining the temporal feature layer, encoding layer, attention layer, vector decomposition layer, and prediction layer into a load prediction model. The LSTM decoder is the core component in the LSTM architecture, used to decode the fixed-length vector generated by the encoder into a variable-length output sequence, suitable for sequence-to-sequence tasks. Monte Carlo Dropout processing is a technique that maintains Dropout activation during the inference phase and samples the prediction results multiple times to quantify model uncertainty, thereby improving the robustness of the model in complex scenarios.

[0030] S26. Based on the load forecasting model, hyperparameter optimization is performed using cross-validation to obtain an optimized forecasting model. Standardized data is input into the optimized forecasting model to obtain multi-time period load forecast sequences and confidence intervals. This invention captures the long-term time dependence characteristics of load through an LSTM network and dynamically weights key time steps using an attention mechanism, significantly improving forecast accuracy by more than 25%. The time series decomposition algorithm breaks down the load into trend, periodic, and random components, enabling the model to learn the changing patterns at different time scales and enhancing its adaptability to complex load patterns. Through multi-layer, multi-modal processing and hyperparameter optimization, the load forecasting model can be made more accurate, improving both forecast accuracy and processing efficiency.

[0031] S3. Based on the multi-period load forecast sequence, voltage safety assessment and key node identification are obtained sequentially through the voltage deviation amplitude coefficient formula, sensitivity coefficient matrix method, and fuzzy comprehensive evaluation method. Existing technologies assess voltage deviation based on single-period load data or simple average load, failing to fully consider the dynamic characteristics of multi-period load changes, resulting in inaccurate voltage deviation amplitude calculations. To address this issue, the specific implementation steps are as follows: S31. Calculate the voltage deviation amplitude coefficient using the voltage deviation amplitude coefficient formula based on the transformer's rated voltage and the multi-period load forecast sequence. The voltage deviation amplitude coefficient is usually expressed as a percentage. It is generally obtained by dividing the absolute value of the difference between the multi-period load forecast sequence and the transformer's rated voltage by the transformer's rated voltage and then multiplying by 100%. It is used to quantify the degree to which the actual voltage deviates from the rated voltage. This will not be elaborated on here.

[0032] S32. Obtain the grid parameters from the transformer's equipment nameplate. Based on the voltage deviation amplitude coefficient and the grid parameters, obtain the node voltage sensitivity matrix using the sensitivity coefficient matrix method. The formula for calculating the node voltage sensitivity matrix is:

[0033] Where S represents the node voltage sensitivity matrix, V represents the node voltage vector, and T represents the transformer turns ratio vector, which is obtained by inverting the Jacobian matrix of the power flow equation. The grid parameters include voltage vector, transformer turns ratio vector, and other related parameters. The sensitivity coefficient matrix method is an analytical method that quantifies the impact of system parameter changes on performance indicators by constructing a matrix. Its core lies in using matrix elements to reveal the complex relationship between parameters and performance, and obtaining sensitivity values ​​through mathematical calculations, such as Taylor expansion and finite difference methods, to provide a theoretical basis for system optimization, fault diagnosis, and robustness assessment. It will not be elaborated on here.

[0034] S33. Based on the voltage deviation amplitude coefficient and the node voltage sensitivity matrix, the voltage safety assessment and key node identification are obtained through the fuzzy comprehensive evaluation method. The fuzzy comprehensive evaluation method is a multi-factor, multi-level comprehensive evaluation method based on fuzzy mathematics theory. It transforms qualitative evaluation into quantitative calculation by introducing fuzzy sets and membership functions. It is suitable for handling complex system evaluation problems with fuzziness and uncertainty. Through a fuzzy rule base and weighted average method, a comprehensive judgment is made, and the safety level is output. The fuzzy comprehensive evaluation method is a commonly used method for fuzzy evaluation, and will not be elaborated upon here. This invention utilizes transformer rated voltage and multi-period load forecast sequences, and calculates the voltage deviation amplitude coefficient using a formula that fully considers load changes over time. This allows for a more accurate calculation of the voltage deviation amplitude coefficient, providing a reliable foundation for subsequent accurate voltage condition assessment. Combining the previously calculated voltage deviation amplitude coefficient, a node voltage sensitivity matrix is ​​constructed using the sensitivity coefficient matrix method. This method closely integrates actual equipment parameters with grid operation data, more accurately reflecting the sensitivity relationship between each node voltage and related factors, improving the understanding of grid voltage characteristics. By comprehensively analyzing the voltage deviation amplitude coefficient and the node voltage sensitivity matrix, the fuzziness and uncertainty of multiple factors are effectively addressed. This enables a comprehensive assessment of voltage safety and accurately identifies key nodes with significant impact on voltage safety, providing a valuable basis for grid operation and maintenance personnel to formulate targeted measures. This helps improve the safety and stability of grid operation and better adapts to the complex and ever-changing operating environment of modern power grids.

[0035] S4. Based on the use of transformers, safety specifications are generated. Within the scope of the safety specifications, based on voltage safety assessment and key node identification, the optimal voltage regulation strategy and strategy confidence are obtained sequentially through a state coding network, a Bayesian deep Q-network, a multi-objective Bellman optimal equation, and multi-criteria decision and confidence. Traditional voltage regulation strategy generation techniques often rely on single-objective optimization or simple rules, making it difficult to handle the complex conflicts between multiple objectives in voltage safety assessment, such as stability, economy, and efficiency. This process often lacks accurate capture of dynamic changes in the system state, resulting in poor strategy adaptability and insufficient consideration of uncertainties, such as load fluctuations and equipment failures, on the impact of voltage regulation. Furthermore, the lack of a quantitative confidence mechanism during strategy evaluation makes it difficult to intuitively judge strategy reliability, and it is prone to safety hazards or economic losses due to over-reliance on experience or simplified models. To solve these problems, the specific implementation steps are as follows: S41. Based on voltage safety assessment and key node identification, a low-dimensional state representation vector is calculated using a state coding network; the formula for calculating the low-dimensional state representation vector is:

[0036] Where s represents the low-dimensional state representation vector, U represents the voltage vector, P and Q represent the power vector and internal energy value, respectively, and level represents the safety level. The variance operation is a commonly used statistical calculation formula, which will not be elaborated here. The state coding network is a neural network structure that converts the state of a system or problem into a low-dimensional representation vector through a specific encoding method. Its core is to map discrete or high-dimensional state information into a compact vector form so that the computer can process it efficiently or optimize the system design. This will not be elaborated here.

[0037] S42. Generate the action interval based on the confidence interval and the transformer's safety specifications. Obtain the posterior distribution of the Q-value using a Bayesian deep Q-network based on the low-dimensional state representation vector and the action interval. The expression for the posterior distribution of the Q-value is:

[0038] in, This means that for each low-dimensional state representation vector s corresponding to an action a in an action interval, the Q value follows a normal distribution. Let represent the mean of the distribution, and be the expected value of Q. The variance of the distribution represents the degree of uncertainty of the Q value. The action range refers to a safe range of power parameters generated by the safety specifications of the transformer, which includes safety thresholds such as operating data, maintenance data, and early warning data. Bayesian deep Q-network is a combination of Bayesian theory and deep Q-network, aiming to improve the decision robustness and interpretability of DQN in complex environments by introducing the uncertainty quantification capability of Bayesian statistics.

[0039] S43. Obtain the current transformer status data. Based on the status data and the posterior distribution of the Q value, calculate the reward vector using a multi-objective reward function. The status data includes the core voltage of the transformer system, sensitivity data, current voltage and current of the equipment, voltage operation adjustment costs, and safety data. This data is the basic data for calculating voltage quality rewards, transformer operation cost rewards, and safety violation penalties. The expression of the multi-objective reward function is:

[0040] Where R represents the reward vector, This indicates a voltage quality reward. This indicates a cost incentive for transformer operation. The formulas for calculating voltage quality rewards, transformer operation cost rewards, and safety violation penalties are commonly used and will not be elaborated here. Voltage quality rewards can be calculated by summing the squares of voltage deviations at all critical nodes. Transformer operation cost rewards are calculated by summing the absolute values ​​of transformer tap changes and all activated or deactivated reactive power compensation equipment, such as capacitor banks, using a weighted summation formula. The weighting coefficients can be obtained through the analytic hierarchy process (AHP). Safety violation penalties are severe, similar to a veto, primarily targeting absolutely prohibited behaviors that could lead to accidents. They are calculated using an indicator function for constraint violations.

[0041] S44. Based on the reward vector and the posterior distribution of Q-values, the Pareto optimal policy set is calculated using the multi-objective Bellman optimal equation. The multi-objective Bellman optimal equation is an extension of multi-objective optimization problems, used to describe how to find the optimal policy and value function through recursion when multiple conflicting or simultaneously optimized objectives exist. It expresses that the optimal value of the current state equals the immediate reward obtained after taking the optimal action in that state, plus the discounted optimal value of the next state. In multi-objective optimization problems, the agent needs to optimize multiple objectives simultaneously, such as time, cost, and safety. These objectives may conflict or require trade-offs. The multi-objective Bellman optimal equation, by extending the form of the Bellman optimal equation, introduces value functions and corresponding weights or priorities for multiple objectives to handle such multi-objective optimization problems. The multi-objective Bellman optimal equation is a commonly used equation to describe the recursive relationship of state-value functions or action-value functions under optimal policies, which will not be elaborated upon here.

[0042] S45. Based on the Pareto optimal policy set, the optimal voltage regulation policy and policy confidence are calculated using multi-criteria decision-making and confidence calculation. The expression for the optimal voltage regulation policy is:

[0043] in, This represents the optimal strategy. The weighting coefficients are obtained through the analytic hierarchy process (AHP). Let the reward vector be represented, and the formula for calculating the policy confidence is:

[0044] in, Indicates policy confidence. Representing state s and optimal policy The standard deviation of the Q-value under the action, This represents the standard deviation of the Q-values ​​under state s and action a. The standard deviation indicates uncertainty; a larger value indicates greater uncertainty in the Q-value estimate for action a. This ratio illustrates the proportion of the uncertainty of the optimal action to the worst-case uncertainty. Multi-criteria decision-making is a method for making decisions under multiple conflicting or simultaneously considered criteria. In optimization problems, when multiple optimization objectives exist, these objectives often cannot be simultaneously optimal. Therefore, a balance or trade-off method is needed to select the most satisfactory solution from multiple feasible solutions. A common implementation of multi-criteria decision-making is the weighted sum method. This method assigns a weight to each objective, combines the value functions of multiple objectives into a comprehensive value function, and then selects the strategy that maximizes the comprehensive value function as the optimal strategy. Confidence calculation is an indicator for evaluating the reliability of decisions or predictions. In optimization problems, confidence can be used to measure the merits or reliability of the selected optimal strategy, i.e., the degree of advantage of the strategy relative to other feasible strategies. Confidence is assessed by calculating the difference in value functions between the optimal and suboptimal strategies, or among all feasible strategies. Specifically, confidence can be calculated using the ratio of the difference between the optimal strategy and the maximum possible value function to all possible actions, which will not be elaborated upon here. This invention dynamically captures system state characteristics through a state-coding network and combines it with a Bayesian deep Q-network to handle uncertainty, achieving dynamic optimization under the multi-objective Bellman optimality equation. It can balance multiple conflicting objectives such as stability and economy, and the multi-criteria decision-making and confidence calculation mechanism further quantifies the strategy's advantages and reliability, providing an intuitive basis for decision-making. This process combines dynamic adaptability, uncertainty robustness, and multi-objective collaborative optimization capabilities, significantly improving the scientific nature and safety of voltage regulation strategies and effectively reducing the risks and costs in transformer regulation caused by the single objective and simplified model of traditional methods.

[0045] S5. Based on the optimal voltage regulation strategy and strategy confidence, specific voltage regulation commands are generated through command mapping. A state prediction model is constructed based on the specific voltage regulation commands. The predicted voltage regulation state output by the state prediction model is used to obtain the voltage qualification rate and voltage regulation effect evaluation through the voltage qualification rate formula. Because existing technologies typically rely on preset rules or shallow models, such as PID control and threshold judgment, for voltage regulation decisions and effect evaluation, the rule settings lack dynamic adaptability and are difficult to cover nonlinear relationships under complex operating conditions. Shallow models have limited feature extraction capabilities and cannot accurately capture the multi-dimensional coupling effect between transformer state and voltage regulation commands, leading to voltage regulation strategies being prone to lag or over-regulation in scenarios such as voltage fluctuations and load changes. The evaluation results also lack robustness due to reliance on a single threshold. To solve these problems, the specific implementation steps are as follows: S51. The optimal voltage regulation strategy and strategy confidence are generated through the instruction mapping method to produce specific voltage regulation instructions. The instruction mapping method is a method that transforms an abstract optimal strategy into specific, executable operation instructions. Its core lies in establishing a clear correspondence between strategy decision variables and actual equipment control instructions. Through preset rules or mathematical models, the abstract strategy output by the optimization algorithm, such as continuous variables and discrete decisions, is converted into physical operation instructions that the equipment can recognize, such as gear adjustment values ​​and switch states. Its essence is a mapping from the strategy space to the execution space, ensuring that the optimization results can directly drive the actual system operation, and the strategy confidence can provide an operating range.

[0046] S52. Obtain the current state of the transformer. Based on the specific voltage regulation command and the current state, construct a state prediction model using a deep neural network. Input the specific voltage regulation command and the current state into the state prediction model to obtain the predicted voltage regulation state. The predicted voltage regulation state includes state-related data such as predicted node voltage vector, predicted branch power flow, predicted total system load, predicted voltage deviation amplitude coefficient, and predicted node voltage sensitivity change. Depending on the actual situation, other state-related data may also be included, which will not be elaborated here. A deep neural network is a prediction tool built on a deep learning architecture. It automatically extracts complex features from the data through multi-layer nonlinear transformations and establishes a mapping relationship between input and output to achieve high-precision prediction tasks. Since deep neural networks are a commonly used prediction model, they will not be elaborated here.

[0047] S53. Obtain the voltage qualification rate using the voltage qualification rate formula based on the predicted voltage regulation status, and generate a voltage regulation effect evaluation based on the voltage qualification rate. If the voltage qualification rate is less than or equal to 95%, the voltage regulation effect is evaluated as unqualified. If the voltage compliance rate is greater than 95%, the voltage regulation effect is evaluated as qualified. The voltage compliance rate formula is one of the core indicators for evaluating the voltage regulation effect. It is used to quantify the proportion of the system voltage maintained within the specified range. For example, if the voltage compliance rate is set to 95%, the voltage compliance rate represents the proportion of the total operating time, that is, 95% of the time is normal voltage, which will not be elaborated here. This invention can flexibly adjust the voltage regulation strategy according to the real-time needs of the system through the command mapping method, avoiding the rigidity of fixed rules. The deep neural network automatically extracts the transformer status, such as voltage, load, and temperature, and the complex nonlinear relationship between them and the voltage regulation command, improving the accuracy of predicting the voltage regulation status. Based on the dynamic threshold of the voltage compliance rate, such as setting it at 95%, the voltage regulation effect is evaluated. This avoids the one-sidedness of a single threshold and enhances the objectivity of the evaluation through a data-driven approach. The voltage regulation strategy is more in line with the actual working conditions, the prediction accuracy is significantly improved, and the evaluation results can guide the iterative optimization of the strategy, forming a closed loop of generation-prediction-evaluation-optimization, which greatly improves the stability and energy efficiency of voltage regulation.

[0048] S6. Based on the voltage qualification rate and voltage regulation effect, synchronization status information is generated through threshold comparison and position difference comparison methods. The final control command is obtained based on multi-condition logic judgment using this synchronization status information. However, when processing transformer voltage regulation control, factors such as voltage qualification rate, voltage regulation effect, and tap position may not be comprehensively considered, leading to inaccurate and incomplete voltage regulation control. This can result in unreasonable overcurrent blocking and asynchronous status information affecting the accuracy of the final control command. To address these issues, the specific implementation steps are as follows: S61. Based on the voltage qualification rate and voltage regulation effect evaluation, the overcurrent blocking result is generated by the threshold comparison method; if the voltage regulation effect evaluation is unqualified, the overcurrent blocking result is blocked; if the voltage regulation effect evaluation is qualified, the overcurrent blocking result is allowed.

[0049] S62. Obtain the position data of multiple taps of the transformer. Based on the position data, synchronize the overcurrent blocking results using the position difference comparison method to obtain synchronization status information. The position difference comparison method is an algorithm that calculates the tap position differences of parallel-running transformers to determine whether they meet the conditions for synchronous operation. Its core is to quantify the deviation of each transformer tap position and compare it with a preset threshold to determine the synchronization status. That is, it traverses all transformer tap position combinations, calculates the absolute difference between each pair, and takes the maximum value as the synchronization status index. This will not be elaborated here.

[0050] S63. The final control command is obtained by using multi-condition logic judgment based on the overcurrent blocking result, synchronization status information, and predicted voltage regulation status. Multi-condition logic judgment is a combination of logic rules based on multiple input flags, such as overcurrent blocking, synchronization status, and predicted status, to determine whether to output a voltage regulation control command. Its core is to ensure that the command is executed only when all safety conditions are met through logical operations such as "AND," "OR," and "NOT." In practice, this step uses "AND" logic for judgment. For example, condition 1: no overcurrent blocking; condition 2: synchronization during single transformer operation or parallel operation; comprehensive judgment: if both condition 1 and condition 2 are met, the final control command is output; otherwise, the command is blocked and an alarm is triggered. This invention generates overcurrent blocking results based on voltage qualification rate and voltage regulation effect evaluation using a threshold comparison method, accurately controlling the overcurrent blocking conditions. It comprehensively and multidimensionally integrates various key factors, making overcurrent blocking more reasonable, status information more accurate and synchronized, and the final control command more precise and reliable, effectively improving the performance and stability of transformer voltage regulation control.

[0051] S7. Obtain the transformer's execution result based on the final control command, update historical data based on the execution result, and perform deep optimization of the state prediction model based on the updated historical data and the current hyperparameters of the state prediction model to obtain a deep state model. Then, obtain the updated control command through the deep state model. Existing technologies heavily rely on precise offline models and expert experience after regulation, making them unable to adapt to dynamic changes in grid operation. Control accuracy and efficiency will significantly decrease. Once the model and parameters are put into operation, they remain fixed, lacking the ability to autonomously learn and optimize from historical data. This results in system performance not improving with operating time and high maintenance costs. To solve these problems, the specific implementation steps are as follows: S71. Complete the adjustment of the transformer according to the final control command and obtain the execution result after adjustment. Generate experience playback data through priority experience playback based on the execution result, and store the experience playback data as updated historical data in the transformer database. S72. Obtain the current hyperparameters of the state prediction model. Based on updated historical data and the current hyperparameters, obtain a dynamic learning rate that can be dynamically adjusted according to the historical gradient magnitude using the AdaGrad adaptive learning rate algorithm. AdaGrad is an adaptive learning rate optimization algorithm designed to solve the problem of low training efficiency or difficulty in convergence caused by a fixed learning rate in traditional gradient descent. It dynamically adjusts the learning rate of each parameter so that the learning rate of frequently updated parameters, such as those with large gradients, gradually decreases, while the learning rate of sparse parameters, such as those with small gradients, remains large, thereby accelerating convergence and improving model performance. This will not be elaborated on here. "Periodic" refers to making dynamic adjustments at intervals to ensure that the final control commands and execution results can achieve the expected results, such as making dynamic adjustments once every 24 hours.

[0052] S73. Based on the dynamic learning rate, a deep state model is obtained by performing deep optimization of the state prediction model through Bayesian deep Q-learning. The deep state model replaces the state prediction model to obtain the updated voltage regulation state, and the updated control command is obtained through S6. This invention, through a closed-loop learning paradigm of priority experience playback – adaptive learning rate – Bayesian deep optimization, can continuously learn from its own operational experience, dynamically adjust decision-making strategies, and quantify decision risks. This continuous learning and optimization significantly improves the system's control accuracy, robustness, and long-term economic efficiency in real-world complex environments.

[0053] Due to the significant limitations of existing technologies in processing complex power system data, and the difficulty in capturing the nonlinear variation law of the coupled influence of multiple factors, this embodiment also proposes a system for an automatic control method for on-load tap change of transformers. This system can significantly improve prediction accuracy, enhance adaptability to complex load patterns, make the load prediction model more accurate, and improve prediction accuracy and processing efficiency.

[0054] Example 2: like Figure 2 As shown, an automatic control system for on-load tap change of a transformer includes a standardization module 1, a load prediction module 2, an evaluation module 3, a strategy module 4, an effect module 5, a control module 6, and a periodic optimization module 7. Standardization Module 1 is used to generate a transformer database based on historical transformer data, and to preprocess the raw data in the transformer database to obtain standardized data. Load forecasting module 2 is used to construct a load forecasting model based on standardized data through long short-term memory network, attention mechanism and seasonal time series decomposition, optimize the load forecasting model to obtain an optimized forecasting model, and input standardized data into the optimized forecasting model to obtain multi-period load forecasting sequences and confidence intervals; Evaluation module 3 is used to obtain voltage safety assessment and key node identification based on the multi-period load forecast sequence through the voltage deviation amplitude coefficient formula, sensitivity coefficient matrix method and fuzzy comprehensive evaluation method. Strategy Module 4 is used to generate safety specifications based on transformer usage. Within the scope of the safety specifications, based on voltage safety assessment and key node identification, the optimal voltage regulation strategy and strategy confidence are obtained sequentially through state coding network, Bayesian deep Q network, multi-objective Bellman optimal equation and multi-criteria decision and confidence. Effect module 5 is used to generate specific voltage regulation commands through command mapping based on the optimal voltage regulation strategy and strategy confidence, construct a state prediction model according to the specific voltage regulation commands, and obtain the voltage qualification rate and voltage regulation effect evaluation from the predicted voltage regulation state output by the state prediction model through the voltage qualification rate formula. Control module 6 is used to generate synchronization status information based on voltage qualification rate and voltage regulation effect through threshold comparison method and position difference comparison method, and to obtain the final control command based on the synchronization status information using multi-condition logic judgment. The periodic optimization module 7 is used to obtain the transformer's execution result based on the final control command, obtain updated historical data based on the execution result, complete the deep optimization of the state prediction model based on the updated historical data and the current hyperparameters of the state prediction model to obtain a deep state model, and obtain updated control commands through the deep state model.

[0055] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0056] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, they are described relatively simply; relevant parts can be referred to the descriptions of the method embodiments.

[0057] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for automatically controlling on-load tap change of a transformer, characterized in that, The method includes the following steps: S1. Generate a transformer database based on the raw data of the transformer, and preprocess the raw data in the transformer database to obtain standardized data; S2. Construct an optimized prediction model based on standardized data, and input the standardized data into the optimized prediction model to obtain multi-period load prediction sequences and confidence intervals; S3. Based on the multi-period load forecast sequence, voltage safety assessment and key node identification are obtained sequentially through the voltage deviation amplitude coefficient formula, sensitivity coefficient matrix method and fuzzy comprehensive evaluation method. S4. Based on the use of transformers, generate safety specifications. Within the scope of the safety specifications, based on voltage safety assessment and key node identification, obtain the optimal voltage regulation strategy and strategy confidence through multi-objective Bellman optimal equation and multi-criteria decision-making. S5. Based on the optimal voltage regulation strategy and strategy confidence, specific voltage regulation commands are generated through command mapping. A state prediction model is constructed according to the specific voltage regulation commands to obtain the predicted voltage regulation state. The voltage qualification rate and voltage regulation effect evaluation are obtained through the voltage qualification rate formula based on the predicted voltage regulation state. S6. Based on the voltage qualification rate and voltage regulation effect, synchronization status information is generated by threshold comparison method and position difference comparison method. Based on the synchronization status information, multi-condition logic judgment is used to obtain the final control command. S7. Obtain the execution result for updating historical data based on the final control instruction, optimize the model based on the updated historical data and the current hyperparameters of the state prediction model to obtain the deep state model, and obtain the update control instruction through the deep state model.

2. The method for on-load tap changing of an automatic control transformer according to claim 1, characterized in that, In step S1, the specific implementation steps are as follows: S11. Based on the raw data in the transformer database, outlier removal is performed using the Raida criterion to obtain qualified data; S12. Based on the qualified data, perform data smoothing and denoising using the wavelet threshold denoising method to obtain denoised data; S13. Obtain standardized data by Z-score standardization based on the denoised data.

3. The method for on-load tap changing of an automatic control transformer according to claim 1, characterized in that, In step S2, the specific implementation steps are as follows: S21. Construct a sliding window with a width of D and a sliding step size of m. Calculate the multi-scale temporal feature matrix of the standardized data within the sliding window using the time series feature extraction formula, and generate the temporal feature layer. S22. Generate a long short-term hidden state sequence using LSTM encoding based on the multi-scale temporal feature matrix, and generate an encoding layer; S23. Obtain the weight distribution through the attention mechanism based on the long and short-term hidden state sequences, obtain the context vector through weighted fusion of the weights based on the weight distribution, and construct the attention layer; S24. Based on the context vector, the vector is decomposed through seasonal time series decomposition to obtain the trend component, periodic component and residual component, and a vector decomposition layer is constructed. S25. Based on the trend component, periodic component, residual component and context vector, the initial multi-period load prediction sequence and initial confidence interval are obtained by sequentially processing the LSTM decoder and Monte Carlo Dropout. The prediction layer is then constructed, and the time series feature layer, encoding layer, attention layer, vector decomposition layer and prediction layer are combined into a load prediction model. S26. Based on the load forecasting model, the hyperparameters are optimized using the cross-validation method to obtain the optimized forecasting model. Standardized data is then input into the optimized forecasting model to obtain multi-period load forecasting sequences and confidence intervals.

4. The method for on-load tap changing of an automatic control transformer according to claim 1, characterized in that, In step S3, the specific implementation steps are as follows: S31. The voltage deviation amplitude coefficient is calculated using the voltage deviation amplitude coefficient formula based on the transformer rated voltage and the multi-period load prediction sequence. S32. Obtain the power grid parameters from the transformer's equipment nameplate, and obtain the node voltage sensitivity matrix using the sensitivity coefficient matrix method based on the voltage deviation amplitude coefficient and the power grid parameters. S33. Based on the voltage deviation amplitude coefficient and the node voltage sensitivity matrix, the voltage safety assessment and key node identification are obtained through the fuzzy comprehensive evaluation method.

5. The method for on-load tap changing of an automatic control transformer according to claim 1, characterized in that, In step S4, the specific implementation steps are as follows: S41. Based on voltage safety assessment and key node identification, a low-dimensional state representation vector is calculated through a state coding network. S42. Generate the action interval based on the confidence interval and the safety specifications of the transformer. Obtain the posterior distribution of the Q value through a Bayesian deep Q network based on the low-dimensional state representation vector and the action interval. S43. Obtain the current state data of the transformer, and calculate the reward vector based on the state data and the posterior distribution of the Q value using a multi-objective reward function. S44. The Pareto optimal policy set is calculated using the multi-objective Bellman optimal equation based on the reward vector and the posterior distribution of the Q-value. S45. Based on the Pareto optimal strategy set, the optimal voltage regulation strategy and strategy confidence are obtained through multi-criteria decision-making and confidence calculation.

6. The method for on-load tap changing of an automatic control transformer according to claim 1, characterized in that, In step S5, the specific implementation steps are as follows: S51. The optimal voltage regulation strategy and strategy confidence are generated by the instruction mapping method to produce specific voltage regulation instructions. S52. Based on the specific voltage regulation command and status data, a status prediction model is constructed through a deep neural network. The specific voltage regulation command and status data are input into the status prediction model to obtain the predicted voltage regulation status. S53. Obtain the voltage qualification rate using the voltage qualification rate formula based on the predicted voltage regulation status, and generate a voltage regulation effect evaluation based on the voltage qualification rate. If the voltage qualification rate is less than or equal to 95%, the voltage regulation effect is evaluated as unqualified. If the voltage qualification rate is greater than 95%, the voltage regulation effect is evaluated as qualified.

7. The method for on-load tap changing of an automatic control transformer according to claim 1, characterized in that, In step S6, the specific implementation steps are as follows: S61. Based on the voltage qualification rate and voltage regulation effect evaluation, the overcurrent blocking result is generated by the threshold comparison method. S62. Obtain the position data of multiple taps of the transformer, and synchronize the overcurrent blocking results according to the position data by using the position difference comparison method to obtain the synchronization status information. S63. The final control command is obtained by using multi-condition logic judgment based on the overcurrent blocking result, synchronization status information and predicted voltage regulation status.

8. The method for on-load tap changing of an automatic control transformer according to claim 1, characterized in that, In step S7, the specific implementation steps are as follows: S71. Complete the adjustment of the transformer according to the final control command and obtain the execution result after adjustment. Generate experience playback data through priority experience playback based on the execution result, and store the experience playback data as updated historical data in the transformer database. S72. Obtain the current hyperparameters of the state prediction model, and obtain a dynamic learning rate that can be dynamically adjusted by the historical gradient magnitude through the AdaGrad adaptive learning rate algorithm based on the updated historical data and the current hyperparameters. S73. Based on the dynamic learning rate, the state prediction model is deeply optimized through Bayesian deep Q-learning to obtain a deep state model. The deep state model replaces the state prediction model and the updated voltage regulation state is obtained. The updated control command is obtained through S6.

9. A system for automatically controlling on-load tap change of a transformer, used in the method for automatically controlling on-load tap change of a transformer as described in any one of claims 1 to 8, characterized in that, The system includes: The standardization module (1) is used to generate a transformer database based on historical transformer data and to preprocess the raw data in the transformer database to obtain standardized data. The load forecasting module (2) is used to construct a load forecasting model based on standardized data through long short-term memory network, attention mechanism and seasonal time series decomposition, optimize the load forecasting model to obtain an optimized forecasting model, and input standardized data into the optimized forecasting model to obtain multi-period load forecasting sequences and confidence intervals; The evaluation module (3) is used to obtain voltage safety assessment and key node identification based on the multi-period load forecast sequence through the voltage deviation amplitude coefficient formula, sensitivity coefficient matrix method and fuzzy comprehensive evaluation method. The strategy module (4) is used to generate safety specifications based on the use of transformers. According to the voltage safety assessment and key node identification, within the scope of the safety specifications, the optimal voltage regulation strategy and strategy confidence are obtained sequentially through state coding network, Bayesian deep Q network, multi-objective Bellman optimal equation and multi-criteria decision and confidence. The effect module (5) is used to generate specific voltage regulation commands through command mapping based on the optimal voltage regulation strategy and strategy confidence, and to build a state prediction model according to the specific voltage regulation commands. The voltage qualification rate and voltage regulation effect evaluation are obtained from the predicted voltage regulation state output by the state prediction model through the voltage qualification rate formula. The control module (6) is used to generate synchronization status information based on voltage qualification rate and voltage regulation effect through threshold comparison method and position difference comparison method, and to obtain the final control command based on the synchronization status information by multi-condition logic judgment. The periodic optimization module (7) is used to obtain the transformer's execution result based on the final control command, obtain updated historical data based on the execution result, complete the deep optimization of the state prediction model based on the updated historical data and the current hyperparameters of the state prediction model to obtain a deep state model, and obtain updated control commands through the deep state model.

Images