A transformer area energy storage system operation regulation method and device

Abstract

The application provides a transformer area energy storage system operation regulation method and device, and belongs to the technical field of power system regulation. The method comprises the following steps: collecting transformer area historical data and constructing a historical database; training a prediction model based on the historical database, inputting real-time data of the transformer area into the prediction model for ultra-short-term prediction, and outputting future power output and load demand prediction values; taking the power output and load demand prediction values as the basis, taking the elimination of line overload, three-phase imbalance and voltage out-of-limit as the optimization target, considering the service life of the energy storage battery and the electricity economy, constructing a multi-objective model and solving the model to obtain an instruction sequence; issuing the instruction sequence to the transformer area energy storage system for execution; collecting real-time data; calculating the execution deviation according to the transformer area actual data, and adaptively updating and correcting the parameters of the prediction model and the multi-objective model. The application solves the transformer area power quality problem, prolongs the service life of the energy storage battery, and reduces the operation cost through intelligent prediction and optimization regulation.

Description

Technical Field

[0001] This application belongs to the field of power system control technology, specifically relating to a method and device for controlling the operation of a distribution area energy storage system. Background Technology

[0002] With the rapid development of new power systems, distributed power sources (such as photovoltaic and wind power) are being connected on a large scale in rural distribution substations. The uncertainty of their output and the volatility of their loads have led to increasingly prominent power quality problems in substation operation, such as line overload, three-phase imbalance, and voltage exceeding limits. Configuring substation energy storage systems and implementing active regulation has become an important solution.

[0003] Currently, energy storage regulation largely relies on strategies based on fixed thresholds or simple rules, such as charging and discharging when voltage exceeds limits or discharging during peak load periods. These methods have significant limitations: First, they suffer from response lag. Threshold triggering mechanisms based on real-time detection cannot predict or proactively adjust for potential future power quality issues. Second, the time-based strategy units are typically designed for single problems (such as voltage or load), lacking the ability to collaboratively optimize multiple objectives such as overload, three-phase imbalance, and voltage deviation. Third, they exhibit poor time adaptability, failing to adapt to dynamic changes in distributed power sources and loads, and struggling to cope with operational characteristic variations caused by external factors such as weather, seasons, and holidays. Finally, they neglect energy storage lifespan; frequent or improper charging and discharging operations accelerate battery aging, and existing strategies often fail to incorporate battery life damage into their optimization objectives.

[0004] In recent years, research on the application of artificial intelligence and big data technologies in power systems has been increasing, such as in load forecasting of regional power grids, power forecasting of renewable energy plants, and operation optimization at the transmission network level. However, when focusing on a single distribution substation at the end of the distribution network, operation and control face drastically different challenges. These substations are small in scale, exhibit significant fluctuations in distributed power sources and loads, and suffer from prominent three-phase imbalances, demanding extremely high real-time, precise, and adaptive control. Existing technological solutions often exhibit limitations when migrating to substation-level scenarios: for example, they struggle to balance the timeliness of real-time decision-making with the systematic coordination of multiple objectives, or they cannot fully adapt to the heterogeneous equipment and fine-grained data operating environment of the substation.

[0005] Therefore, there is an urgent need for a distribution network energy storage operation and control method that can integrate multi-source data, achieve ultra-short-term forecasting, and complete multi-objective dynamic optimization under multiple constraints, so as to improve the safety, economy and power quality of distribution network operation. Summary of the Invention

[0006] In a first aspect, embodiments of this application provide a method for the operation and control of a distribution area energy storage system, comprising the following steps: S1. Collect and store historical operation data of the transformer area, build a historical database, mine and analyze the data in the historical database, identify the inherent defects and optimization space in the operation of the transformer area, and output the defect analysis results; S2. Train a deep learning prediction model based on historical databases, and input the real-time operation data of the transformer area into the trained deep learning prediction model to make ultra-short-term predictions, and output the predicted output value of distributed power sources and the predicted load demand value for a future set time period. S3. Based on the predicted output value of distributed power sources and the predicted load demand value, with the optimization goal of alleviating or eliminating line overload, three-phase imbalance and voltage over-limit problems, and considering the life damage of energy storage batteries and the economic efficiency of electricity use, a multi-objective optimization model is constructed and solved to obtain the optimal charging and discharging power command sequence of the energy storage system in each phase. S4. Send the optimal charging and discharging power command sequence to the energy storage system in the distribution area for execution, so as to balance the supply and demand in the distribution area, correct the three-phase current, and support the voltage; at the same time, collect and monitor the actual operating status data of the distribution area after the execution of the optimal charging and discharging power command sequence in real time. S5. Based on the actual operating status data of the transformer area, calculate the strategy execution deviation, and based on the strategy execution deviation and the newly accumulated operating data, adaptively update and correct the parameters of the deep learning prediction model and the parameters of the multi-objective optimization model.

[0007] Furthermore, the specific steps of step S1 are as follows: S11. Continuously collect multi-source data from the metering terminal and sensing terminal on the transformer area side. The multi-source data includes historical and real-time load data for each phase, distributed power generation data, energy storage system status data, three-phase current and voltage data, meteorological data, and calendar information. S12. Clean, align, and normalize the collected raw data, and store it in a pre-built historical database; S13. Based on the historical database, clustering algorithms are used to classify historical operating periods, identify typical problem scenarios that lead to line overload, severe three-phase imbalance or voltage over-limit, and extract feature sets of each typical problem scenario to generate historical operating condition analysis results. S14. Based on the historical operating condition analysis results, quantitatively assess the degree to which the regulation potential of the energy storage system has not been utilized during historical operation, and output the defect analysis results including the characteristic set of typical problem scenarios and the historical regulation potential assessment.

[0008] Furthermore, the specific steps of step S2 are as follows: S21. Construct a deep learning prediction model and train it using time-series data from a historical database until the model converges. S22. At each prediction time, construct the real-time feature vector as input. The real-time feature vector Includes load sequences for each phase over n historical time periods. Distributed power generation output sequence Meteorological data series and the calendar features of the current moment. ; S23. Transfer the real-time feature vector The input is fed into a trained deep learning prediction model, which outputs predictions for the next k time periods, including predicted load demand values ​​for each phase. and the predicted output value of distributed power sources .

[0009] Furthermore, in step S21, the deep learning prediction model is a Long Short-Term Memory (LSTM) network model; The structure of the Long Short-Term Memory network model includes an input layer, at least one LSTM layer, and an output layer connected in sequence. The calculation process for a single LSTM layer at time t is as follows: Forgotten Gate:

[0010] Input Gate:

[0011] Candidate memory units:

[0012] Memory unit update:

[0013] Output gate:

[0014] Hidden state output:

[0015] in, This is a temporal slice of the real-time feature vector that serves as input; This is the hidden state from the previous moment; This represents the state of the memory unit from the previous moment; , , , These are the weight matrices for the forget gate, input gate, candidate memory units, and output gate, respectively. , , , These are the bias vectors for the forget gate, input gate, candidate memory unit, and output gate, respectively. Use the Sigmoid activation function; The training process of a deep learning prediction model is as follows: using the feature vectors from previous historical time steps... and the corresponding actual load and power generation data As training samples, mean squared error is used as the loss function. Backpropagation algorithm is used to optimize the weight matrix and bias vector through gradient descent until the loss function converges to a preset threshold. The formula for the loss function is as follows:

[0016] Where m is the number of samples in the current training batch; The true value of the j-th sample is taken from the historical load or historical power generation data in the historical database. This is the output of the deep learning prediction model for the j-th sample.

[0017] Furthermore, the deep learning prediction model in step S21 is a gated recurrent unit network (GRU) model; The structure of the gated recurrent unit network model includes an input layer, at least one GRU layer, and an output layer connected in sequence. The calculation process for a single GRU layer at time t is as follows: Update Gate:

[0018] Reset Door:

[0019] Candidate hidden states:

[0020] Final hidden state output:

[0021] in, This is a temporal slice of the real-time feature vector that serves as input; This is the hidden state from the previous moment; , , These are the weight matrices for the update gate, reset gate, and candidate hidden states, respectively. , , These are the bias vectors for the update gate, reset gate, and candidate hidden states, respectively. Use the Sigmoid activation function; The training process of a deep learning prediction model is as follows: using the feature vectors from previous historical time steps... and the corresponding actual load and power generation data As training samples, mean squared error is used as the loss function. The weight matrix and bias vector are optimized by gradient descent using the backpropagation algorithm until the loss function converges to a preset threshold. The formula for the loss function is as follows:

[0022] Where m is the number of samples in the current training batch; The true value of the j-th sample is taken from the historical load or historical power generation data in the historical database. This is the output of the deep learning prediction model for the j-th sample.

[0023] Furthermore, the specific steps of step S3 are as follows: S31. Construct the following multi-objective optimization function with the core of alleviating line overload, three-phase imbalance and voltage over-limit, while taking into account energy storage life loss and electricity economy. At the same time, define the upper and lower limits of the energy storage system's state of charge (SOC), the charging and discharging power limit, and the system power balance as constraints.

[0024] in, This is a line overload penalty term, the value of which is proportional to the square of the line current exceeding the safety threshold; This is a penalty term for three-phase unbalance, and its value is proportional to the absolute value of the three-phase current unbalance. This is a voltage deviation penalty term, the value of which is proportional to the square of the deviation of the node voltage from the rated voltage range; This is a penalty term for battery life loss, and its value is related to the absolute value of the current charge / discharge power and the rate of change of SOC. , , , The initial values ​​are set based on the defect analysis results, representing the weighting coefficients for each objective item. S32. Perform periodic solutions according to a preset decision cycle. At the beginning of each decision cycle, using the prediction result of S2 as input, solve the multi-objective optimization function to obtain the optimal charging and discharging power command sequence of the energy storage system in each phase covering the preset future control period. And it is distributed to the energy storage system.

[0025] Furthermore, the specific steps of step S4 are as follows: S41. The local controller of the energy storage system receives a sequence of commands for optimal charging and discharging power. S42. The energy storage system performs corresponding charging or discharging operations in each phase according to the optimal charging and discharging power command sequence, and outputs or absorbs active and reactive power. S43. While executing the optimal charging and discharging power command sequence, the real-time operating data of the transformer area is collected in real time through the metering terminal. The real-time operating data includes three-phase voltage, current, power, and actual output and SOC data of the energy storage system. S44. Upload the collected actual operating status data in real time.

[0026] Furthermore, the specific steps of step S5 are as follows: S51. Collect actual operating status data and compare it with the key indicators of voltage, current and power expected by the multi-objective optimization function in step S31, and calculate the deviation of each indicator as the strategy execution deviation. S52. Store the continuous prediction inputs, optimization instructions and actual results within the current set time period as a data group in the rolling time window for use in updating the deep learning prediction model; S53. Periodically use new data within the rolling time window to incrementally train the deep learning prediction model and update the corresponding network weights; S54. Based on policy execution bias, dynamically adjust the weight coefficients of the multi-objective optimization function using gradient-based methods or reinforcement learning algorithms. , , , This allows the optimization strategy to continuously adapt to changes in the operating characteristics of the transformer area.

[0027] Furthermore, the specific steps in step S54 of dynamically adjusting the weight coefficients of the multi-objective optimization function using a gradient-based method are as follows: S541. At each update time of the multi-objective optimization function, calculate the total objective function value of the multi-objective optimization function. Relative to each weight coefficient gradient The total objective function value For the current scrolling time window All line overload penalties actually generated during all decision-making cycles within the mouth Three-phase imbalance penalty item Voltage deviation penalty item Battery life loss penalty item The weighted sum of , i.e. ; S542. Based on the calculated gradient, update the weight coefficients of the multi-objective optimization function according to the following formula:

[0028] in, This is the new weight coefficient for the i-th term; Let be the old weight coefficient of the i-th term; The learning rate is used to control the update step size. The regularization coefficient is used. Initial weights set based on the results of the step defect analysis; S543. Normalize the updated weight coefficients so that the sum of the weight coefficients is a fixed constant; S544. Write the normalized weight coefficients back into the multi-objective optimization function for solving subsequent decision cycles.

[0029] Secondly, embodiments of this application also provide an operation control device for a transformer substation energy storage system, comprising: The data acquisition and historical analysis module is used to collect and store historical operating data of the transformer area, build a historical database, mine and analyze the data in the historical database, identify the inherent defects and optimization space in the operation of the transformer area, and output the defect analysis results. The ultra-short-term forecasting module is used to train a deep learning forecasting model based on historical databases, and input the real-time operation data of the transformer area into the trained deep learning forecasting model to make ultra-short-term forecasts, and output the forecast values ​​of distributed power output and load demand for a set period of time in the future. The multi-objective optimization decision module is used to construct and solve a multi-objective optimization model based on the output forecast of distributed power sources and the load demand forecast, with the optimization objectives of alleviating or eliminating line overload, three-phase imbalance and voltage over-limit problems, and taking into account the life damage of energy storage batteries and the economic efficiency of electricity use, so as to obtain the optimal charging and discharging power command sequence of the energy storage system in each phase. The instruction execution and monitoring module is used to send the optimal charging and discharging power instruction sequence to the energy storage system in the distribution area for execution, so as to balance the supply and demand in the distribution area, correct the three-phase current, and support the voltage; at the same time, it collects and monitors the actual operating status data of the distribution area after the optimal charging and discharging power instruction sequence is executed in real time. The closed-loop feedback and correction module is used to calculate the strategy execution deviation based on the actual operating status data of the transformer area, and to adaptively update and correct the parameters of the deep learning prediction model and the multi-objective optimization model based on the strategy execution deviation and the newly accumulated operating data.

[0030] As can be seen from the above technical solutions, this application has the following advantages: The method and device for operation and control of the transformer substation energy storage system provided in this application effectively solve problems such as line overload, three-phase imbalance, and voltage exceeding limits through deep learning prediction and multi-objective optimization, maximizing the comprehensive benefits and power quality of the transformer substation operation; it achieves real-time, precise, and self-correcting control of the transformer substation energy storage equipment, automatically adapting to changes in the operating characteristics of the transformer substation; it considers the lifespan loss of energy storage batteries during the optimization process, avoids improper charging and discharging operations, extends battery life, and reduces operating costs; it takes into account the economics of electricity consumption, optimizes electricity costs, and improves control efficiency, reducing the impact of power quality problems on the power grid and users; based on multi-source data fusion and historical operating condition analysis, it provides a scientific basis for optimization strategies, improving the accuracy and reliability of decision-making. Attached Figure Description

[0031] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the operation and control method of the transformer substation energy storage system of the present invention.

[0033] Figure 2 This is a schematic diagram of the operation and control device for the transformer substation energy storage system of the present invention. Detailed Implementation

[0034] The various embodiments of this disclosure will be described more fully in the detailed steps of the operation and control method for the energy storage system in the following text. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0035] This embodiment provides a method for the operation and control of a transformer substation energy storage system. By utilizing deep learning and multi-objective optimization techniques, it can accurately predict load and power output, effectively alleviating problems such as line overload, three-phase imbalance, and voltage exceeding limits.

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figure 1 The diagram shows a flowchart of a specific embodiment of a method for operating and controlling an energy storage system in a distribution area. The method includes the following steps: S1. Collect and store historical operation data of the transformer area, build a historical database, mine and analyze the data in the historical database, identify the inherent defects and optimization space in the operation of the transformer area, and output the defect analysis results; It should be noted that comprehensively collecting historical operational data of the transformer substations and constructing a historical database provides data resources for subsequent analysis and forecasting, ensuring the data foundation for control strategies; through the mining and analysis of historical data, the inherent defects and optimization space in the operation of the transformer substations can be accurately identified, providing a reasonable basis for setting optimization targets; S2. Train a deep learning prediction model based on historical databases, and input the real-time operation data of the transformer area into the trained deep learning prediction model to make ultra-short-term predictions, and output the predicted output value of distributed power sources and the predicted load demand value for a future set time period. It should be noted that training deep learning prediction models based on historical databases fully utilizes information from historical data to improve the model's predictive capabilities. Using the trained model for ultra-short-term predictions can help anticipate changes in distributed power output and load demand, providing key inputs for regulation and enhancing the proactiveness of regulation. S3. Based on the predicted output value of distributed power sources and the predicted load demand value, with the optimization goal of alleviating or eliminating line overload, three-phase imbalance and voltage over-limit problems, and considering the life damage of energy storage batteries and the economic efficiency of electricity use, a multi-objective optimization model is constructed and solved to obtain the optimal charging and discharging power command sequence of the energy storage system in each phase. It should be noted that, in order to solve the power quality problem and take into account both the lifespan of energy storage batteries and the economics of electricity consumption, a comprehensive multi-objective optimization model is constructed to ensure the comprehensiveness and balance of the control strategy; the solution is executed according to the preset decision cycle to quickly generate the optimal charging and discharging power command sequence covering future control periods, thereby improving the timeliness and effectiveness of control. S4. Send the optimal charging and discharging power command sequence to the energy storage system in the distribution area for execution, so as to balance the supply and demand in the distribution area, correct the three-phase current, and support the voltage; at the same time, collect and monitor the actual operating status data of the distribution area after the execution of the optimal charging and discharging power command sequence in real time. It should be noted that the optimal charging and discharging power command sequence is sent to the energy storage system for execution, enabling the energy storage system to actively output or absorb power, thereby achieving supply and demand balance, three-phase current correction, and voltage support in the distribution area, directly improving the operating status of the distribution area. During the execution process, the actual operating status data of the distribution area is collected in real time, providing real-time data support for subsequent feedback correction and ensuring accurate evaluation of the control effect. S5. Based on the actual operating status data of the transformer area, calculate the strategy execution deviation, and based on the strategy execution deviation and the newly accumulated operating data, adaptively update and correct the parameters of the deep learning prediction model and the parameters of the multi-objective optimization model. It should be noted that by comparing the actual operating status data with the optimization target, the strategy execution deviation is calculated, and the relevant data is stored in the rolling time window to provide a data basis for model updates. Based on the deviation and new data, the prediction model and optimization model are adaptively updated and corrected, so that the control strategy can continuously adapt to the changes in the operating characteristics of the transformer area and maintain the stability and superiority of the control effect.

[0038] This embodiment achieves intelligent control of the energy storage system in the transformer substation by collecting multi-source data, using deep learning prediction and multi-objective optimization, effectively solving power quality problems, balancing energy storage life and economy, and improving operational efficiency.

[0039] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process in this embodiment, another method for the operation and control of a distribution substation energy storage system is provided. Taking a rural 10kV distribution substation as an example, this substation covers 120 residential users, 3 distributed photovoltaic power stations (total installed capacity 500kW), and 1 set of substation energy storage system (rated capacity 200kWh, rated power 100kW). The substation has typical problems: the output of distributed photovoltaic power fluctuates drastically due to the influence of sunlight, which can easily cause line overload during peak hours; the three-phase distribution of residential electricity load is uneven, with the three-phase imbalance reaching up to 25%; when the peak electricity consumption coincides with the low photovoltaic output, the terminal voltage deviation exceeds the rated value by ±7%. This method can achieve precise management of the power quality problems in the substation, while taking into account the lifespan of the energy storage battery and the economic efficiency of electricity use. Smart metering and sensing terminals are installed on the low-voltage side of the transformer in the distribution area, at each distributed photovoltaic outlet, energy storage system interface, and three-phase outgoing branch. The data collection frequency is 1 minute / time, covering load data of each phase, distributed power generation data, energy storage system SOC and power data, and three-phase voltage and current data. Small weather stations are deployed near the distribution area to collect meteorological data such as wind speed, light intensity, temperature, and humidity at a frequency of 5 minutes / time. The communication module adopts 4G / LoRa dual-mode communication. The metering terminal and the sensing terminal transmit the collected data to the smart energy terminal through 4G / LoRa. The smart energy terminal realizes data interaction and command transmission with the cloud energy storage platform through the APN gateway. The communication latency is controlled within 500ms. The energy storage system is equipped with a three-phase independently controlled energy storage converter (PCS), which supports independent adjustment of active and reactive power, with a charging and discharging power adjustment range of 0-100kW and a SOC operating range of 20%-80%. It has the ability to receive and execute charging and discharging power commands for each phase sent from the cloud. The cloud-based energy storage platform, deployed on the office network, integrates modules for data acquisition and historical analysis, ultra-short-term forecasting, multi-objective optimization decision-making, command execution and monitoring, and closed-loop feedback and correction. It possesses functions for data storage, model training, optimization solving, command issuance, and status monitoring. The method includes the following steps: S1. Collect and store historical operational data of the transformer area, construct a historical database, perform data mining and analysis on the historical database, identify inherent defects and optimization opportunities in the operation of the transformer area, and output the defect analysis results; the specific steps of step S1 are as follows: S11. Continuously collect multi-source data from the metering terminal and sensing terminal on the transformer area side. The multi-source data includes historical and real-time load data for each phase, distributed power generation data, energy storage system status data, three-phase current and voltage data, meteorological data, and calendar information. For example, multi-source data is continuously collected from metering terminals, sensing terminals and weather stations on the transformer substation side, including historical load data for each phase from January 1, 2023 to December 31, 2023, distributed photovoltaic power generation data, SOC and charge / discharge power data of energy storage system, three-phase voltage and current data, as well as corresponding meteorological data (light intensity, temperature, humidity, etc.) and calendar information (weekdays / holidays, seasons, etc.). S12. Clean, align, and normalize the collected raw data, and store it in a pre-built historical database; For example, the collected raw data is preprocessed. Outliers caused by equipment failure (such as sudden voltage spikes and drops) are removed, and missing data is filled using linear interpolation. Data from different collection frequencies are aligned to a 1-minute time scale. Load, power generation, meteorological parameters, and other data are normalized (normalization range is [0,1]). The processed data is stored in a historical database, which uses a distributed storage architecture to ensure data security and access efficiency. S13. Based on the historical database, clustering algorithms are used to classify historical operating periods, identify typical problem scenarios that lead to line overload, severe three-phase imbalance or voltage over-limit, and extract feature sets of each typical problem scenario to generate historical operating condition analysis results. For example, based on a historical database, the K-means clustering algorithm is used to classify historical operating periods. With 5 clusters, the similarity of indicators such as line current, three-phase imbalance, and voltage deviation across different time periods is calculated to identify 3 typical problem scenarios: Scenario 1: Peak photovoltaic output (sunlight intensity ≥ 800W / m²), line current exceeds safety threshold by 15%, duration concentrated between 11:00-14:00; Scenario 2: Severe three-phase imbalance, mainly occurring during peak electricity consumption from 18:00 to 21:00, where the load of phase A is 2.3 times that of phase C, and the three-phase imbalance is ≥20%. Scenario 3: Voltage exceeding limit scenario, which mostly occurs between 21:00 and 22:00, when photovoltaic output drops sharply and the load is high, and the terminal voltage is lower than 93% of the rated voltage; Extract feature sets for each scenario. For example, the features of scenario 1 are high light intensity, high photovoltaic output, and excessive line current, and generate historical operating condition analysis results. S14. Based on the historical operating condition analysis results, quantitatively assess the degree to which the regulation potential of the energy storage system has not been utilized during historical operation, and output the defect analysis results including the characteristic set of typical problem scenarios and the historical regulation potential assessment. For example, based on historical operating condition analysis results, the regulation potential of the energy storage system is quantitatively assessed. Calculations show that in typical problem scenarios, the energy storage system only utilizes 60% of its regulation potential, indicating significant untapped capacity. The output defect analysis results include three typical problem scenario feature sets, historical regulation potential assessment (40% untapped potential), and the probability of occurrence for each scenario (Scenario 1: 25%, Scenario 2: 30%, Scenario 3: 20%). S2. Train a deep learning prediction model based on historical databases, and input real-time operating data of the distribution area into the trained deep learning prediction model for ultra-short-term prediction, outputting the predicted output value of distributed power sources and the predicted load demand value for a future set time period; the specific steps of step S2 are as follows: S21. Construct a deep learning prediction model and train it using time-series data from a historical database until the model converges. In step S21, the deep learning prediction model is the Long Short-Term Memory (LSTM) network model; The structure of the Long Short-Term Memory network model includes an input layer, at least one LSTM layer, and an output layer connected in sequence. The calculation process for a single LSTM layer at time t is as follows: Forgotten Gate:

[0040] Input Gate:

[0041] Candidate memory units:

[0042] Memory unit update:

[0043] Output gate:

[0044] Hidden state output:

[0045] in, This is a temporal slice of the real-time feature vector that serves as input; This is the hidden state from the previous moment; This represents the state of the memory unit from the previous moment; , , , These are the weight matrices for the forget gate, input gate, candidate memory units, and output gate, respectively. , , , These are the bias vectors for the forget gate, input gate, candidate memory unit, and output gate, respectively. Use the Sigmoid activation function; The training process of a deep learning prediction model is as follows: using the feature vectors from previous historical time steps... and the corresponding actual load and power generation data As training samples, mean squared error is used as the loss function. Backpropagation algorithm is used to optimize the weight matrix and bias vector through gradient descent until the loss function converges to a preset threshold. The formula for the loss function is as follows:

[0046] Where m is the number of samples in the current training batch; The true value of the j-th sample is taken from the historical load or historical power generation data in the historical database. This is the output of the deep learning prediction model for the j-th sample. For example, a Long Short-Term Memory (LSTM) network model is selected as the deep learning prediction model. The model structure includes an input layer, two LSTM layers (each with 64 neurons), and an output layer. The training set is time-series data from January 1, 2023 to November 30, 2023, from a historical database. Input features include load sequences for each phase, distributed power generation output sequences, and meteorological data sequences for 60 historical time periods (1 hour). The output is the predicted load demand for each phase and the predicted output of distributed power generation for the next 12 time periods (12 minutes). Mean squared error is used as the loss function. The backpropagation algorithm is used to optimize the weight matrix and bias vector through gradient descent. The learning rate is set to 0.001. After 1000 training iterations, the loss function converges to a preset threshold of 0.005, and the model training is complete. S22. At each prediction time, construct the real-time feature vector as input. The real-time feature vector Includes load sequences for each phase over n historical time periods. Distributed power generation output sequence Meteorological data series and the calendar features of the current moment. ; For example, at the prediction time of 9:00 on January 1, 2024, a real-time feature vector is constructed, which includes the A, B, and C three-phase load sequences, photovoltaic power output sequences, light intensity / temperature / humidity sequences for 60 time periods (8:00-9:00) before 9:00, as well as the calendar features (weekday, winter) at the current time. S23. Transfer the real-time feature vector The input is fed into a trained deep learning prediction model, which outputs predictions for the next k time periods, including predicted load demand values ​​for each phase. and the predicted output value of distributed power sources ; For example, the real-time feature vector is input into the trained LSTM model, which outputs the prediction results for the next 12 time periods (9:01-9:12). For instance, it is predicted that the photovoltaic output will remain at around 450kW from 9:05 to 9:08, with phase A load at 80kW, phase B load at 65kW, and phase C load at 35kW, showing a clear three-phase load imbalance. S3. Based on the predicted output and load demand of distributed power sources, and with the optimization objectives of alleviating or eliminating line overload, three-phase imbalance, and voltage exceeding limits, while considering the lifespan and energy economy of energy storage batteries, a multi-objective optimization model is constructed and solved to obtain the optimal charging and discharging power command sequence for each phase of the energy storage system; the specific steps of step S3 are as follows: S31. Construct the following multi-objective optimization function with the core of alleviating line overload, three-phase imbalance and voltage over-limit, while taking into account energy storage life loss and electricity economy. At the same time, define the upper and lower limits of the energy storage system's state of charge (SOC), the charging and discharging power limit, and the system power balance as constraints.

[0047] in, This is a line overload penalty term, the value of which is proportional to the square of the line current exceeding the safety threshold; This is a penalty term for three-phase unbalance, and its value is proportional to the absolute value of the three-phase current unbalance. This is a voltage deviation penalty term, the value of which is proportional to the square of the deviation of the node voltage from the rated voltage range; This is a penalty term for battery life loss, and its value is related to the absolute value of the current charge / discharge power and the rate of change of SOC. , , , The initial values ​​are set based on the defect analysis results, representing the weighting coefficients for each objective item. For example, a multi-objective optimization function is constructed with the core objective of alleviating line overload, three-phase imbalance and voltage over-limit, while taking into account energy storage life loss and electricity consumption economy. Line overload penalty Set the line safety current threshold to 120A. ( The actual current of the line during time period t. =0.05), the value is proportional to the square of the line current exceeding the safety threshold; Three-phase imbalance penalty : These are the currents of phases A, B, and C, respectively. =0.03, the value is proportional to the absolute value of the three-phase current unbalance; Voltage deviation penalty The rated voltage is set at 220V, with an allowable deviation range of ±7% (204.6V-235.4V). for( Let t be the node voltage during time period t. =220V, =0.04), the value is proportional to the square of the deviation of the node voltage from the rated voltage range; Battery life loss penalty : ( For energy storage charging and discharging power, The rate of change of SOC =0.02, =0.5), the value is related to the absolute value of the current charge / discharge power and the rate of change of SOC; the initial value of the weighting coefficient is set based on the defect analysis results. =0.3 (Line overload issues account for 25%) =0.35 (three-phase imbalance accounts for 30%) =0.25 (voltage exceeding limits account for 20%) =0.1 (energy storage lifetime weight), satisfying ; The constraints include: upper and lower limits of energy storage SOC (20%-80%), charging and discharging power limits (-100kW to 100kW, negative values ​​are for charging and positive values ​​are for discharging), and system power balance (distributed power output + energy storage output = load demand). S32. Perform periodic solutions according to a preset decision cycle. At the beginning of each decision cycle, using the prediction result of S2 as input, solve the multi-objective optimization function to obtain the optimal charging and discharging power command sequence of the energy storage system in each phase covering the preset future control period. And distribute it to the energy storage system; For example, the decision cycle is set to 12 minutes. At the beginning of each decision cycle (e.g., 9:00), the prediction result of S2 is used as input, and the particle swarm optimization algorithm is used to solve the multi-objective optimization function. The particle swarm size is set to 50, and the number of iterations is 50. The optimal charging and discharging power command sequence for each phase of the energy storage system in the next 12 time periods (9:01-9:12) is obtained. For example, the command for the time period 9:05-9:08 is: phase A discharges 10kW, phase B discharges 5kW, and phase C charges 20kW. The load distribution is balanced by the three-phase differentiated charging and discharging, while avoiding line overload. S4. The optimal charging and discharging power command sequence is sent to the energy storage system in the distribution area for execution to balance the supply and demand in the distribution area, correct the three-phase current, and support the voltage; at the same time, the actual operating status data of the distribution area after the execution of the optimal charging and discharging power command sequence is collected and monitored in real time; the specific steps of step S4 are as follows: S41. The local controller of the energy storage system receives a sequence of commands for optimal charging and discharging power. S42. The energy storage system performs corresponding charging or discharging operations in each phase according to the optimal charging and discharging power command sequence, and outputs or absorbs active and reactive power. S43. While executing the optimal charging and discharging power command sequence, the real-time operating data of the transformer area is collected in real time through the metering terminal. The real-time operating data includes three-phase voltage, current, power, and actual output and SOC data of the energy storage system. S44. Upload the collected actual operating status data in real time; For example, the local controller of the energy storage system receives the optimal charging and discharging power command sequence sent from the cloud via 4G / Lora communication; According to the command sequence, the energy storage system performs charging and discharging operations on each phase through the three-phase independently controlled PCS. During the period from 9:05 to 9:08, the A phase PCS outputs 10kW of active power, the B phase outputs 5kW of active power, and the C phase absorbs 20kW of active power, so as to achieve precise adjustment of active and reactive power, balance the supply and demand of the distribution area, and correct the three-phase current. While executing instructions, the metering terminal collects real-time operating data of the transformer area, including three-phase voltage, current, power, and actual output and SOC data of the energy storage system, with a collection frequency of 1 minute / time. The collected actual operating status data is uploaded to the cloud energy storage platform in real time via 4G / Lora for subsequent deviation calculation and model correction; S5. Based on the actual operating status data of the transformer area, calculate the strategy execution deviation, and based on the strategy execution deviation and the newly accumulated operating data, adaptively update and correct the parameters of the deep learning prediction model and the parameters of the multi-objective optimization model; the specific steps of step S5 are as follows: S51. Collect actual operating status data and compare it with the key indicators such as voltage, current, and power expected by the multi-objective optimization function in step S31, and calculate the deviation of each indicator as the strategy execution deviation. For example, the actual operating status data collected between 9:01 and 9:12 is compared with the key indicators expected by the multi-objective optimization function; for example, the actual A-phase current is 85kW, the B-phase current is 70kW, and the C-phase current is 55kW, the three-phase imbalance is reduced to 8%, which is lower than the predicted deviation of 5%; the actual line current is 110A, which does not exceed the safety threshold, and the voltage is maintained between 215V and 225V; the deviation of each indicator is calculated, and the line overload deviation is 0, the three-phase imbalance deviation is 3%, the voltage deviation is 0, and the battery life loss deviation is 0.01, which are taken as the strategy execution deviation; S52. Store the continuous prediction inputs, optimization instructions and actual results within the current set time period as a data group in the rolling time window for use in updating the deep learning prediction model; For example, the predicted input, optimization instructions, and actual results for the period from 9:00 to 9:12 are stored as a data set in a rolling time window (the window size is set to 24 hours) for updating the deep learning prediction model; S53. Periodically use new data within the rolling time window to incrementally train the deep learning prediction model and update the corresponding network weights; For example, at 2:00 AM every day (the off-peak period for electricity consumption in the distribution area), new data (data from the last 24 hours) within the rolling time window is used to incrementally train the LSTM prediction model, the learning rate is adjusted to 0.0005, and the model network weights are updated to adapt the model to the changes in the operating characteristics of the distribution area. S54. Based on policy execution bias, dynamically adjust the weight coefficients of the multi-objective optimization function using gradient-based methods or reinforcement learning algorithms. , , , This allows the optimization strategy to continuously adapt to changes in the operating characteristics of the transformer area; The specific steps in step S54 of dynamically adjusting the weight coefficients of the multi-objective optimization function using a gradient-based method are as follows: S541. At each update time of the multi-objective optimization function, calculate the total objective function value of the multi-objective optimization function. Relative to each weight coefficient gradient The total objective function value For the current scrolling time window All line overload penalties actually generated during all decision-making cycles within the mouth Three-phase imbalance penalty item Voltage deviation penalty item Battery life loss penalty item The weighted sum of , i.e. ; S542. Based on the calculated gradient, update the weight coefficients of the multi-objective optimization function according to the following formula:

[0048] in, This is the new weight coefficient for the i-th term; Let be the old weight coefficient of the i-th term; The learning rate is used to control the update step size. The regularization coefficient is used. Initial weights set based on the results of the step defect analysis; S543. Normalize the updated weight coefficients so that the sum of the weight coefficients is a fixed constant; S544. Write the normalized weight coefficients back into the multi-objective optimization function for solving subsequent decision cycles; For example, based on policy execution bias, a gradient-based method is used to dynamically adjust the weight coefficients of the multi-objective optimization function; At each update time (2:30 AM daily), calculate the total objective function value. Relative to the gradient of each weight coefficient, the overall objective function value is the weighted sum of all penalty terms actually generated in all decision cycles within the current rolling time window; based on the gradient calculation results, according to the formula... Update the weight coefficients, where =0.01 (learning rate) =0.05 (regularization coefficient) As initial weights; for example, because the actual improvement in three-phase imbalance is better than expected, the gradient calculation results show... It needs to be lowered after the update. It becomes 0.32; Normalize the updated weight coefficients to ensure After normalization =0.31、 =0.32、 =0.26、 =0.11; The normalized new weight coefficients are written back into the multi-objective optimization function for use in solving subsequent decision cycles.

[0049] In one embodiment of the present invention, unlike the above embodiments, the deep learning prediction model in step S21 is a gated recurrent unit network (GRU) model. The structure of the gated recurrent unit network model includes an input layer, at least one GRU layer, and an output layer connected in sequence. The calculation process for a single GRU layer at time t is as follows: Update Gate:

[0050] Reset Door:

[0051] Candidate hidden states:

[0052] Final hidden state output:

[0053] in, This is a temporal slice of the real-time feature vector that serves as input; This is the hidden state from the previous moment; , , These are the weight matrices for the update gate, reset gate, and candidate hidden states, respectively. , , These are the bias vectors for the update gate, reset gate, and candidate hidden states, respectively. Use the Sigmoid activation function; The training process of a deep learning prediction model is as follows: using the feature vectors from previous historical time steps... and the corresponding actual load and power generation data As training samples, mean squared error is used as the loss function. The weight matrix and bias vector are optimized by gradient descent using the backpropagation algorithm until the loss function converges to a preset threshold. The formula for the loss function is as follows:

[0054] Where m is the number of samples in the current training batch; The true value of the j-th sample is taken from the historical load or historical power generation data in the historical database. This is the output of the deep learning prediction model for the j-th sample.

[0055] In one embodiment of the present invention, unlike the above embodiments, the specific steps of dynamically adjusting the weight coefficients of the multi-objective optimization function using a reinforcement learning algorithm in step S54 are as follows: S541. The energy storage control system of the distribution area is modeled as a Markov decision process, wherein: state Defined as the current operating characteristics of the transformer area, it includes at least the load of each phase, the output of distributed power sources, the state of charge of the energy storage system, the three-phase voltage and current deviation, and historical control effect evaluation indicators. action Defined as the weight coefficients of a multi-objective optimization model The adjustment vector, ; award Defined as the negative comprehensive penalty term that actually occurs in the next decision cycle, i.e. Maximizing it is equivalent to minimizing the overall objective function value in actual operation; S542. A deep deterministic policy gradient algorithm is used to construct an actor-commentator network framework; Action Network by state The input determines the output action (weighting coefficient adjustment). ; Critics Network by state and actions Input, output state-action value function The estimated value; S543. The agent executes actions in each decision cycle. After adjusting the weights and applying them to the optimization model, observe the new state of the environmental feedback. and rewards Transfer samples Store to the experience replay buffer; Periodically sample a batch of samples from the buffer and update the commentator network parameters by minimizing the temporal difference error. :

[0056] in, For the target Q value, As a discount factor, Batch size; Then, the actor network parameters are updated through deterministic policy gradient. To maximize expected return:

[0057] S544. The trained actor network As an online strategy, at each model update time, based on the current state... Generate weight coefficient adjustment action ; According to the action Update the current weights: The updated weight coefficients are then normalized. The normalized new weight coefficients Write back to the multi-objective optimization model in step S32.

[0058] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0059] like Figure 2 As shown, the following are embodiments of the operation and control device for the distribution area energy storage system provided in this disclosure. This system and the operation and control method for the distribution area energy storage system in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the operation and control device for the distribution area energy storage system, please refer to the embodiments of the operation and control method for the distribution area energy storage system described above.

[0060] The device includes: The data acquisition and historical analysis module is used to collect and store historical operating data of the transformer area, build a historical database, mine and analyze the data in the historical database, identify the inherent defects and optimization space in the operation of the transformer area, and output the defect analysis results. The ultra-short-term forecasting module is used to train a deep learning forecasting model based on historical databases, and input the real-time operation data of the transformer area into the trained deep learning forecasting model to make ultra-short-term forecasts, and output the forecast values ​​of distributed power output and load demand for a set period of time in the future. The multi-objective optimization decision module is used to construct and solve a multi-objective optimization model based on the output forecast of distributed power sources and the load demand forecast, with the optimization objectives of alleviating or eliminating line overload, three-phase imbalance and voltage over-limit problems, and taking into account the life damage of energy storage batteries and the economic efficiency of electricity use, so as to obtain the optimal charging and discharging power command sequence of the energy storage system in each phase. The instruction execution and monitoring module is used to send the optimal charging and discharging power instruction sequence to the energy storage system in the distribution area for execution, so as to balance the supply and demand in the distribution area, correct the three-phase current, and support the voltage; at the same time, it collects and monitors the actual operating status data of the distribution area after the optimal charging and discharging power instruction sequence is executed in real time. The closed-loop feedback and correction module is used to calculate the strategy execution deviation based on the actual operating status data of the transformer area, and to adaptively update and correct the parameters of the deep learning prediction model and the multi-objective optimization model based on the strategy execution deviation and the newly accumulated operating data.

[0061] This embodiment achieves precise management of power quality issues in transformer substations through the interactive collaboration of data acquisition and historical analysis modules, ultra-short-term prediction modules, multi-objective optimization decision-making modules, instruction execution and monitoring modules, and closed-loop feedback and correction modules. This extends the lifespan of energy storage batteries, reduces electricity costs, and improves the reliability and economy of system operation.

[0062] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for operation and control of a transformer substation energy storage system, characterized in that, Includes the following steps: S1. Collect and store historical operation data of the transformer area, build a historical database, mine and analyze the data in the historical database, identify the inherent defects and optimization space in the operation of the transformer area, and output the defect analysis results; S2. Train a deep learning prediction model based on historical databases, and input the real-time operation data of the transformer area into the trained deep learning prediction model to make ultra-short-term predictions, and output the predicted output value of distributed power sources and the predicted load demand value for a future set time period. S3. Based on the predicted output value of distributed power sources and the predicted load demand value, with the optimization goal of alleviating or eliminating line overload, three-phase imbalance and voltage over-limit problems, and considering the life damage of energy storage batteries and the economic efficiency of electricity use, a multi-objective optimization model is constructed and solved to obtain the optimal charging and discharging power command sequence of the energy storage system in each phase. S4. Send the optimal charging and discharging power command sequence to the energy storage system in the distribution area for execution, so as to balance the supply and demand in the distribution area, correct the three-phase current, and support the voltage; at the same time, collect and monitor the actual operating status data of the distribution area after the execution of the optimal charging and discharging power command sequence in real time. S5. Based on the actual operating status data of the transformer area, calculate the strategy execution deviation, and based on the strategy execution deviation and the newly accumulated operating data, adaptively update and correct the parameters of the deep learning prediction model and the parameters of the multi-objective optimization model.

2. The method for operation and control of a distribution area energy storage system according to claim 1, characterized in that, The specific steps of step S1 are as follows: S11. Continuously collect multi-source data from the metering terminal and sensing terminal on the transformer area side. The multi-source data includes historical and real-time load data for each phase, distributed power generation data, energy storage system status data, three-phase current and voltage data, meteorological data, and calendar information. S12. Clean, align, and normalize the collected raw data, and store it in a pre-built historical database; S13. Based on the historical database, clustering algorithms are used to classify historical operating periods, identify typical problem scenarios that lead to line overload, severe three-phase imbalance or voltage over-limit, and extract feature sets of each typical problem scenario to generate historical operating condition analysis results. S14. Based on the historical operating condition analysis results, quantitatively assess the degree to which the regulation potential of the energy storage system has not been utilized during historical operation, and output the defect analysis results including the characteristic set of typical problem scenarios and the historical regulation potential assessment.

3. The method for operation and control of a distribution area energy storage system according to claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Construct a deep learning prediction model and train it using time-series data from a historical database until the model converges. S22. At each prediction time, construct the real-time feature vector as input. The real-time feature vector Includes load sequences for each phase over n historical time periods. Distributed power generation output sequence Meteorological data series and the calendar features of the current moment. ; S23. Transfer the real-time feature vector The input is fed into a trained deep learning prediction model, which outputs predictions for the next k time periods, including predicted load demand values ​​for each phase. and the predicted output value of distributed power sources .

4. The method for operation and control of a distribution area energy storage system according to claim 3, characterized in that, In step S21, the deep learning prediction model is the Long Short-Term Memory (LSTM) network model; The structure of the Long Short-Term Memory network model includes an input layer, at least one LSTM layer, and an output layer connected in sequence. The calculation process for a single LSTM layer at time t is as follows: Forgotten Gate: Input Gate: Candidate memory units: Memory unit update: Output gate: Hidden state output: in, This is a temporal slice of the real-time feature vector that serves as input; This is the hidden state from the previous moment; This represents the state of the memory unit from the previous moment; , , , These are the weight matrices for the forget gate, input gate, candidate memory units, and output gate, respectively. , , , These are the bias vectors for the forget gate, input gate, candidate memory unit, and output gate, respectively. Use the Sigmoid activation function; The training process of a deep learning prediction model is as follows: using the feature vectors from previous historical time steps... and the corresponding actual load and power generation data As training samples, mean squared error is used as the loss function. Backpropagation algorithm is used to optimize the weight matrix and bias vector through gradient descent until the loss function converges to a preset threshold. The formula for the loss function is as follows: Where m is the number of samples in the current training batch; The true value of the j-th sample is taken from the historical load or historical power generation data in the historical database. This is the output of the deep learning prediction model for the j-th sample.

5. The method for operation and control of a distribution area energy storage system according to claim 3, characterized in that, The deep learning prediction model in step S21 is a gated recurrent unit network (GRU) model. The structure of the gated recurrent unit network model includes an input layer, at least one GRU layer, and an output layer connected in sequence. The calculation process for a single GRU layer at time t is as follows: Update Gate: Reset Door: Candidate hidden states: Final hidden state output: in, This is a temporal slice of the real-time feature vector that serves as input; This is the hidden state from the previous moment; , , These are the weight matrices for the update gate, reset gate, and candidate hidden states, respectively. , , These are the bias vectors for the update gate, reset gate, and candidate hidden states, respectively. Use the Sigmoid activation function; The training process of a deep learning prediction model is as follows: using the feature vectors from previous historical time steps... and the corresponding actual load and power generation data As training samples, mean squared error is used as the loss function. The weight matrix and bias vector are optimized by gradient descent using the backpropagation algorithm until the loss function converges to a preset threshold. The formula for the loss function is as follows: Where m is the number of samples in the current training batch; The true value of the j-th sample is taken from the historical load or historical power generation data in the historical database. This is the output of the deep learning prediction model for the j-th sample.

6. The method for operation and control of a distribution area energy storage system according to claim 1, characterized in that, The specific steps of step S3 are as follows: S31. Construct the following multi-objective optimization function with the core of alleviating line overload, three-phase imbalance and voltage over-limit, while taking into account energy storage life loss and electricity economy. At the same time, define the upper and lower limits of the energy storage system's state of charge (SOC), the charging and discharging power limit, and the system power balance as constraints. in, This is a line overload penalty term, the value of which is proportional to the square of the line current exceeding the safety threshold; This is a penalty term for three-phase unbalance, and its value is proportional to the absolute value of the three-phase current unbalance. This is a voltage deviation penalty term, the value of which is proportional to the square of the node voltage deviation from the rated voltage range; This is a penalty term for battery life loss, and its value is related to the absolute value of the current charge / discharge power and the rate of change of SOC. , , , The initial values ​​are set based on the defect analysis results, representing the weighting coefficients for each objective item. S32. Perform periodic solutions according to a preset decision cycle. At the beginning of each decision cycle, using the prediction result of S2 as input, solve the multi-objective optimization function to obtain the optimal charging and discharging power command sequence of the energy storage system in each phase covering the preset future control period. And it is distributed to the energy storage system.

7. The method for operation and control of a distribution area energy storage system according to claim 1, characterized in that, The specific steps of step S4 are as follows: S41. The local controller of the energy storage system receives a sequence of commands for optimal charging and discharging power. S42. The energy storage system performs corresponding charging or discharging operations in each phase according to the optimal charging and discharging power command sequence, and outputs or absorbs active and reactive power. S43. While executing the optimal charging and discharging power command sequence, the real-time operating data of the transformer area is collected in real time through the metering terminal. The real-time operating data includes three-phase voltage, current, power, and actual output and SOC data of the energy storage system. S44. Upload the collected actual operating status data in real time.

8. The method for operation and control of a distribution area energy storage system according to claim 1, characterized in that, The specific steps of step S5 are as follows: S51. Collect actual operating status data and compare it with the key indicators of voltage, current and power expected by the multi-objective optimization function in step S31, and calculate the deviation of each indicator as the strategy execution deviation. S52. Store the continuous prediction inputs, optimization instructions and actual results within the current set time period as a data group in the rolling time window for use in updating the deep learning prediction model; S53. Periodically use new data within the rolling time window to incrementally train the deep learning prediction model and update the corresponding network weights; S54. Based on policy execution bias, dynamically adjust the weight coefficients of the multi-objective optimization function using gradient-based methods or reinforcement learning algorithms. , , , This allows the optimization strategy to continuously adapt to changes in the operating characteristics of the transformer area.

9. The method for operation and control of a distribution area energy storage system according to claim 8, characterized in that, The specific steps in step S54 of dynamically adjusting the weight coefficients of the multi-objective optimization function using a gradient-based method are as follows: S541. At each update time of the multi-objective optimization function, calculate the total objective function value of the multi-objective optimization function. Relative to each weight coefficient gradient The total objective function value For the current scrolling time window All line overload penalties actually generated during all decision-making cycles within the mouth Three-phase imbalance penalty item Voltage deviation penalty item Battery life loss penalty item The weighted sum of , i.e. ; S542. Based on the calculated gradient, update the weight coefficients of the multi-objective optimization function according to the following formula: in, Let i be the new weight coefficient for the i-th term; Let be the old weight coefficient of the i-th term; The learning rate is used to control the update step size. The regularization coefficient is used. Initial weights set based on the results of the step defect analysis; S543. Normalize the updated weight coefficients so that the sum of the weight coefficients is a fixed constant; S544. Write the normalized weight coefficients back into the multi-objective optimization function for solving subsequent decision cycles.

10. A control device for the operation of a transformer substation energy storage system, characterized in that, include: The data acquisition and historical analysis module is used to collect and store historical operating data of the transformer area, build a historical database, mine and analyze the data in the historical database, identify the inherent defects and optimization space in the operation of the transformer area, and output the defect analysis results. The ultra-short-term forecasting module is used to train a deep learning forecasting model based on historical databases, and input the real-time operation data of the transformer area into the trained deep learning forecasting model to make ultra-short-term forecasts, and output the forecast values ​​of distributed power output and load demand for a set period of time in the future. The multi-objective optimization decision module is used to construct and solve a multi-objective optimization model based on the output forecast of distributed power sources and the load demand forecast, with the optimization objectives of alleviating or eliminating line overload, three-phase imbalance and voltage over-limit problems, and taking into account the life damage of energy storage batteries and the economic efficiency of electricity use, so as to obtain the optimal charging and discharging power command sequence of the energy storage system in each phase. The instruction execution and monitoring module is used to send the optimal charging and discharging power instruction sequence to the energy storage system in the distribution area for execution, so as to balance the supply and demand in the distribution area, correct the three-phase current, and support the voltage; at the same time, it collects and monitors the actual operating status data of the distribution area after the optimal charging and discharging power instruction sequence is executed in real time. The closed-loop feedback and correction module is used to calculate the strategy execution deviation based on the actual operating status data of the transformer area, and to adaptively update and correct the parameters of the deep learning prediction model and the multi-objective optimization model based on the strategy execution deviation and the newly accumulated operating data.

Images