Hybrid energy storage system optimization scheduling method based on AI intelligent regulation and control

Abstract

The invention discloses a hybrid energy storage system optimization scheduling method based on AI intelligent regulation and control, and relates to the technical field of hybrid energy storage, and the method comprises the following steps: collecting real-time data, and carrying out the cross verification of data consistency through a multi-source data fusion technology; and identifying data abnormity caused by sensor faults, communication delay and environmental interference by combining adaptive threshold detection with a statistical analysis method, and eliminating abnormal data. According to the method, energy storage scheduling is optimized through data cleaning, time sequence prediction and reinforcement learning, and the intelligent management and dynamic adaptive capacity is improved. Multi-source data fusion and anomaly detection are adopted to ensure data accuracy, energy consumption is predicted by means of LSTM and Transform, and an energy storage strategy is optimized in advance. By combining reinforcement learning and system dynamic adjustment charging and discharging, the photovoltaic consumption rate is improved, the electricity purchasing cost is reduced, the SOC is intelligently controlled to be 40%-80%, and the service life of the battery is prolonged. Meanwhile, the system stability is improved through anomaly detection and correction, and the maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of hybrid energy storage, and particularly to an optimized scheduling method for a hybrid energy storage system based on AI intelligent regulation. Background Art

[0002] The optimized scheduling of a hybrid energy storage system based on AI intelligent regulation refers to the use of artificial intelligence technology to intelligently optimize the management of the hybrid energy storage system (usually including battery energy storage and supercapacitors, etc.) in a photovoltaic energy storage base station, so as to improve energy utilization efficiency, reduce operating costs and enhance system stability. Through AI algorithms (such as machine learning, deep reinforcement learning, etc.), it analyzes photovoltaic power generation, load demand, grid electricity price and energy storage device status in real time, and makes intelligent decisions on when to store energy, when to discharge, and how to reasonably allocate energy between different energy storage media, so as to achieve the optimal scheduling of energy. This technology can effectively suppress the volatility of photovoltaic power generation, reduce the abandonment of light, improve the reliability of base station power supply, and reduce electricity costs in peak-valley electricity price scheduling.

[0003] The prior art has the following deficiencies: In the process of optimizing the scheduling of a hybrid energy storage system based on AI intelligent regulation, the failure of the scheduling strategy caused by data noise and outliers may bring serious consequences. The energy storage system relies on sensors to collect data such as photovoltaic power generation, load demand, and battery status in real time. If sensor failures, communication delays or environmental interference cause data anomalies, the AI scheduling algorithm may misidentify the system state, and then execute incorrect energy allocation strategies. For example, misjudging the remaining capacity of the battery may lead to overcharging or over-discharging, affecting the battery life and even causing safety accidents. In addition, if the AI model does not cover enough abnormal situations during the training process, the scheduling system may lose robustness when facing sudden anomalies, resulting in unstable power supply of the base station and affecting the normal operation of the communication network.

[0004] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure, and therefore it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The object of the present invention is to provide an optimized scheduling method for a hybrid energy storage system based on AI intelligent regulation, which optimizes energy storage scheduling through data cleaning, time series prediction and reinforcement learning, enabling the system to have intelligent management and dynamic adaptation capabilities. By adopting multi-source data fusion and anomaly detection technologies, the accuracy of data such as photovoltaic power generation, load demand, and battery SOC is ensured. With the help of LSTM and Transformer, future energy consumption is predicted, energy storage strategies are optimized in advance, and power supply instability is prevented. Combining reinforcement learning, the system can dynamically adjust the charge and discharge plan, improve the photovoltaic accommodation rate, reduce the grid power purchase cost, and intelligently control the SOC within the range of 40%-80% to extend the battery life. At the same time, the anomaly scenario detection and correction mechanism can cope with extreme load fluctuations, battery aging and abnormal environmental temperatures, improve system stability and reduce long-term maintenance costs, so as to solve the problems in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: An optimized scheduling method for a hybrid energy storage system based on AI intelligent regulation, comprising the following steps:

[0007] Collect real-time data and cross-verify data consistency through multi-source data fusion technology; use adaptive threshold detection combined with statistical analysis methods to identify data anomalies caused by sensor failures, communication delays and environmental interferences, and eliminate the abnormal data;

[0008] Adopt a fusion algorithm based on historical data regression analysis and physical modeling to correct the marked abnormal data, and use adjacent time window data for interpolation compensation to reduce the impact of data missing on scheduling decisions;

[0009] Use a time series prediction model to predict the photovoltaic power generation, load demand and energy storage state in the future time, establish a scheduling early warning mechanism, and adjust the energy distribution strategy in advance when sudden anomalies are predicted to reduce the impact of anomalies on the stability of the energy storage system;

[0010] Combine the reinforcement learning algorithm, construct a reward function, and optimize the energy scheduling strategy based on the distributed reinforcement learning framework to achieve dynamic optimal scheduling of the energy storage system under different working conditions;

[0011] During the execution of AI scheduling, combine confidence estimation and Bayesian optimization methods to dynamically monitor the reliability of scheduling decisions. When the confidence level is lower than the set threshold, trigger a rule correction mechanism based on an expert system to reasonably adjust the scheduling results and ensure robustness in abnormal situations;

[0012] Adopt a federated learning architecture to enable multiple base stations to share abnormal data patterns and collaboratively optimize the AI model on the premise of protecting data privacy; dynamically expand the training set through an active learning mechanism to improve the adaptability of the model to data noise and outliers, and improve the robustness of the overall scheduling system.

[0013] Preferably, real-time data is collected, and the data consistency is cross-validated through multi-source data fusion technology; the specific steps of using adaptive threshold detection combined with statistical analysis methods to identify data anomalies caused by sensor failures, communication delays, and environmental interferences, and eliminating the abnormal data are as follows:

[0014] Collect key data of photovoltaic, battery, and load, and perform time alignment, format standardization, and interpolation processing to ensure data consistency and availability;

[0015] Adopt multi-source data fusion technology to cross-validate data from different sensors and improve data accuracy;

[0016] Combine adaptive threshold detection, statistical analysis, and time series prediction methods to dynamically identify data anomalies and improve the sensitivity and adaptability of anomaly detection;

[0017] Fill in the abnormal data through sliding window filtering, regression analysis, and deep learning prediction, and optimize the compensation strategy in combination with physical constraints to ensure the stable operation of the energy storage system.

[0018] Preferably, a fusion algorithm based on historical data regression analysis and physical modeling is adopted to correct the marked abnormal data, and interpolation compensation is performed using data in adjacent time windows to reduce the impact of data loss on scheduling decisions. The specific steps are as follows:

[0019] Classify the abnormal data, analyze its impact on the scheduling strategy, and select the optimal compensation strategy to reduce the negative impact brought by data loss;

[0020] Use the regression model to predict and correct the abnormal data to make it conform to the historical trend and improve the accuracy and reliability of the data;

[0021] Conduct rationality constraints on the regression-corrected data through physical modeling to ensure that the corrected data conforms to physical laws and avoid the impact of incorrect compensation on system operation;

[0022] Adopt linear interpolation, spline interpolation, and Kalman filtering methods to fill in short-term missing data, and optimize the compensation in combination with physical constraints to ensure data continuity and scheduling stability.

[0023] Preferably, a time series prediction model is used to predict the photovoltaic power generation, load demand, and energy storage state in the future time, establish a scheduling early warning mechanism, and adjust the energy distribution strategy in advance when sudden anomalies are predicted to reduce the impact of anomalies on the stability of the energy storage system. The specific steps are as follows:

[0024] Sort out and normalize the data of photovoltaic power generation, load demand, and energy storage state, and adopt time window technology to capture long-term and short-term trends to ensure stable data quality;

[0025] Select a time series prediction model for training, optimize the hyperparameters and perform cross-validation to improve the prediction accuracy of photovoltaic power generation, load demand, and energy storage status;

[0026] Set an early warning threshold based on the prediction results, dynamically adjust the energy distribution before the decline of photovoltaic power generation or the surge of load, and optimize the scheduling strategy by combining reinforcement learning;

[0027] Optimize the energy storage scheduling strategy before an anomaly occurs, and adopt a model adaptive adjustment mechanism to correct the prediction error, improving the stability and prediction ability of the energy storage system.

[0028] Preferably, combine the reinforcement learning algorithm, construct a reward function, and optimize the energy scheduling strategy based on the distributed reinforcement learning framework. The specific steps for the energy storage system to achieve dynamic optimal scheduling under different working conditions are as follows;

[0029] Combine the time series prediction results to define the state space of reinforcement learning to construct a dynamic scheduling environment;

[0030] Design a reward function by comprehensively considering economy, battery life, load stability, and renewable energy utilization rate, guiding the agent to learn the optimal energy scheduling strategy, and dynamically adjusting the weights to adapt to different application scenarios;

[0031] Adopt reinforcement learning algorithms such as PPO / DDPG, and use the distributed reinforcement learning framework to accelerate training, enabling the agent to learn the optimal energy scheduling strategy under different working conditions, improving the learning efficiency and convergence speed;

[0032] Through online learning and the model adaptive adjustment mechanism, enable the reinforcement learning agent to continuously optimize the scheduling strategy, dynamically adapt to environmental changes and sudden anomalies, and improve the stability and interpretability of the energy storage system.

[0033] Preferably, during the execution of the AI scheduling, combine the confidence estimation and Bayesian optimization method to dynamically monitor the reliability of the scheduling decision. When the confidence is lower than the set threshold, trigger a rule correction mechanism based on the expert system to reasonably adjust the scheduling result to ensure robustness in case of anomalies. The specific steps are as follows:

[0034] During the execution of the Al scheduling, first evaluate the confidence of the current scheduling strategy to judge the reliability of the decision. Calculate the mean and variance of the scheduling strategy through multiple forward propagations to quantify the uncertainty of the scheduling strategy. The formula is as follows:

[0035]

[0036] Where C t is the confidence of the scheduling decision, σ[P t is the uncertainty of the current scheduling decision, E[Pt ] is the expected value of the scheduling decision at the current moment, ∈ is a very small positive number to prevent the denominator from approaching zero;

[0037] When the confidence level C t Below the set threshold C thresh When , the rule correction mechanism is triggered to ensure the stability of the scheduling strategy;

[0038] When the confidence level is lower than the threshold C thresh When the current dispatching parameters are optimized to find the optimal dispatching value, the dispatching strategy is dynamically adjusted using Bayesian optimization, so that the optimized decision reduces the risk and improves the economy and stability of the energy storage system. The objective function of Bayesian optimization is defined as follows:

[0039]

[0040] In the formula, is the optimized battery charging and discharging power, R(P) is the reward function, which takes into account factors such as grid electricity price, battery life, load fluctuation, etc., λ is the risk weight coefficient, which controls the degree of penalty for uncertainty. is the value of P that maximizes the objective function.

[0041] Preferably, although Bayesian optimization provides optimized scheduling parameters However, it may still lead to unacceptable scheduling decisions in extreme cases. Therefore, a rule correction mechanism based on an expert system is adopted to correct Perform final correction to ensure the safety and feasibility of the scheduling strategy. The rule correction formula is expressed as follows:

[0042]

[0043] In the formula, is the final battery charge and discharge power, I(·) is the indicator function, which takes 1 if the condition is met, otherwise it takes 0, α is the smoothing weight factor, ΔP smooth It is a smooth adjustment value based on load fluctuation.

[0044] Preferably, a federated learning architecture is adopted to enable multiple base stations to share abnormal data patterns and collaboratively optimize the AI ​​model under the premise of protecting data privacy; the training set is dynamically expanded through an active learning mechanism to improve the model's adaptability to data noise and outliers, and the specific steps to improve the robustness of the overall scheduling system are as follows:

[0045] In the process of optimizing the scheduling of the hybrid energy storage system, the data of a single base station is limited, and the abnormal data patterns vary due to various factors such as geographical location, climate, and load characteristics. Therefore, to improve the generalization ability of the Al scheduling model, a federated learning architecture is adopted to enable multiple base stations to share abnormal data patterns while ensuring that data privacy is not leaked. The core calculation formula for aggregating the federated model is as follows:

[0046]

[0047] In the formula, θ (j+1) is the global model parameter in the (j + 1)-th round, N is the total number of base stations, w i is the data volume of base station i, E is the normalization factor of the total data volume of all base stations, is the model parameter of base station i during the j-th round of training, γ is the weight coefficient of abnormal data for model update, M is the normalization factor of the total abnormal data volume of all base stations, is the model gradient of base station i after local training;

[0048] Since the adaptability of the energy storage system scheduling AI model to abnormal data depends on the diversity and coverage of the training data, after optimizing the global model through federated learning, an active learning mechanism is further adopted to dynamically expand the training set and improve the robustness of the model to noise and outliers. The sample selection of active learning is based on the principle of maximizing information gain, and the calculation formula is as follows:

[0049]

[0050] In the formula, S * is the selected subset of active learning samples, S c is the candidate data set, which contains all samples to be screened. x represents a certain data point in the candidate sample set S c . p(y|x, θ (j+1) ) is the predicted probability distribution of sample x under the global model parameter θ (j+1) . H(p(y|x, θ (j+1) )) is the prediction entropy of the current global model. K is the number of sub-models used to calculate uncertainty, represents the parameter of the k-th sub-model after the (j + 1)-th round of training, is to maximize the information gain and select valuable samples.

[0051] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0052] Through data cleaning, time series prediction, and reinforcement learning to optimize the scheduling strategy, the energy storage system can achieve intelligent management in complex environments and enhance its adaptability to dynamic changes. First, multi-source data fusion and anomaly detection technologies are adopted to ensure the accuracy of key data such as photovoltaic power generation, load demand, and battery SOC, avoiding problems such as sensor failures and communication delays from affecting scheduling decisions. Secondly, with the help of time series prediction models such as LSTM and Transformer, the system can predict future photovoltaic power generation and load demand in advance, and establish a scheduling warning mechanism to optimize the energy storage strategy before anomalies occur, preventing power supply instability caused by sudden load shocks. In addition, by combining reinforcement learning to optimize the scheduling strategy, the system can autonomously learn the optimal scheduling plan according to grid electricity prices, photovoltaic output, and load fluctuations. For example, reduce discharging during peak electricity prices to reduce battery losses while ensuring stable power supply. Through intelligent optimization, the system can dynamically adjust the strategy under different climate conditions, load patterns, and energy storage states, enhancing the self-adaptability of the energy storage system and improving the long-term stable operation ability of the photovoltaic energy storage base station.

[0053] Through intelligent control, the system can comprehensively consider factors such as grid electricity prices, photovoltaic power generation prediction, and battery life, and dynamically adjust the charge and discharge plan, thereby reducing operating costs. For example, the system can preferentially charge during off-peak electricity price periods and release energy storage during peak electricity prices, reducing grid power purchase expenditures while increasing the local consumption rate of photovoltaic power generation. In addition, to prevent the battery from aging due to frequent deep charge and discharge, the system optimizes the scheduling through reinforcement learning, intelligently controls the battery SOC within the optimal working range of 40%-80%, reduces overcharging and over-discharging, and improves the battery health (SOH). At the same time, combined with the abnormal scenario detection and intelligent correction mechanism, when detecting extreme load fluctuations, battery aging, or abnormal environmental temperature, the system can automatically adjust the scheduling strategy to prevent the battery from overheating or discharging abnormally, improving battery life and reducing long-term operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0055] Figure 1 It is a method flow chart of the optimization scheduling method for a hybrid energy storage system based on AI intelligent control of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] Example embodiments will now be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the example embodiments to those skilled in the art.

[0057] The present invention provides an optimized scheduling method for a hybrid energy storage system based on AI intelligent regulation as Figure 1 shown, comprising the following steps:

[0058] Collect real-time data such as photovoltaic power generation, load demand, battery status, etc., and cross-verify the data consistency through multi-source data fusion technology; use adaptive threshold detection combined with statistical analysis methods to identify data anomalies caused by sensor failures, communication delays, and environmental interference, and eliminate the abnormal data;

[0059] The specific steps of collecting real-time data such as photovoltaic power generation, load demand, battery status, etc., and cross-verify the data consistency through multi-source data fusion technology; use adaptive threshold detection combined with statistical analysis methods to identify data anomalies caused by sensor failures, communication delays, and environmental interference, and eliminate the abnormal data are as follows:

[0060] Collect key data of photovoltaic, battery, and load, and perform time alignment, format standardization, and interpolation processing to ensure data consistency and availability;

[0061] In the process of optimizing the scheduling of the hybrid energy storage system with AI intelligent regulation, it is first necessary to collect key parameters such as photovoltaic power generation, base station load demand, and battery state of charge (SOC). This process involves multiple sensors, such as photovoltaic current / voltage sensors, current / temperature sensors in the battery management system (BMS), smart meters, etc., and uploads the data to the cloud or local edge computing device through wireless or wired communication methods (such as LoRa, NB-IoT, Modbus). To ensure data integrity, the system needs to preprocess the collected raw data, including timestamp alignment, unit conversion, data format standardization, etc. For example, if the data refresh rates of some sensors are different, the system needs to perform interpolation processing so that all data is comparable at the same time point.

[0062] Adopt multi-source data fusion technology to cross-verify the data of different sensors, and use methods such as redundant measurement and Kalman filtering to improve data accuracy;

[0063] Since the energy storage system involves multiple independent data sources, to ensure data consistency, multi-source data fusion technology needs to be adopted to cross-verify the data from different sensors. For example, redundant sensor data can be used for comparison, such as obtaining the power generation data of the PV MPPT controller and the electricity meter simultaneously and calculating whether the error range between the two is within the set threshold. For different measurement methods of the same variable (such as the battery SOC can be calculated by the BMS or estimated by the coulomb integration method), Kalman filtering or Bayesian estimation methods can be used for fusion to obtain a more accurate state estimate. In addition, during the data fusion process, historical trend analysis can be introduced to detect whether the current data conforms to a reasonable change pattern, thereby improving the data quality.

[0064] Combined with adaptive threshold detection, statistical analysis and time series prediction methods, dynamically identify data anomalies and improve the sensitivity and adaptability of anomaly detection;

[0065] After multi-source data fusion, the system needs to use adaptive threshold detection combined with statistical analysis methods to identify and mark data anomalies. The adaptive threshold detection method can be based on the mean and standard deviation of the statistical data in a sliding window to dynamically adjust the threshold to adapt to the data characteristics in different environments. For example, during normal operation, the voltage fluctuation range of the battery is small, while in extreme weather (such as high temperature or low temperature), the battery performance may have a short-term anomaly. Therefore, the system can adjust the anomaly detection threshold according to the real-time environment. In addition, use box plot analysis (IQR method) or Z-score method to calculate outliers. If a data point deviates from the mean by more than a certain standard deviation, it can be determined as abnormal data. At the same time, use time series analysis techniques (such as ARIMA or LSTM) to predict the normal data range and compare it with the actual data to further enhance the accuracy of anomaly identification.

[0066] Fill in the abnormal data through sliding window filtering, regression analysis and deep learning prediction, and optimize the compensation strategy in combination with physical constraints to ensure the stable operation of the energy storage system;

[0067] For the detected abnormal data, the system needs to adopt appropriate elimination and compensation strategies. First, for short-term anomalies (such as voltage mutations caused by instantaneous jitter of sensors), the sliding window mean filtering method can be used for smoothing; for long-term anomalies (such as continuous anomalies of data caused by sensor failures), the data point needs to be eliminated and historical data is used for intelligent compensation. Intelligent compensation can combine physical modeling and machine learning methods. For example, fill in the missing values through regression models or deep learning (such as Transformer prediction), and at the same time ensure that the compensated data meets the physical constraint conditions (such as the battery SOC cannot exceed 100%). In addition, the system can send an alarm to the operation and maintenance personnel when an anomaly occurs and record the characteristics of the abnormal data to optimize future anomaly detection strategies and improve the overall stability of the system.

[0068] Adopt a fusion algorithm based on historical data regression analysis and physical modeling to correct the marked abnormal data, and use the data in adjacent time windows for interpolation compensation to reduce the impact of data missing on scheduling decisions;

[0069] The specific steps of adopting a fusion algorithm based on historical data regression analysis and physical modeling to correct the marked abnormal data, and using the data in adjacent time windows for interpolation compensation to reduce the impact of data missing on scheduling decisions are as follows:

[0070] Classify the abnormal data, analyze its impact on the scheduling strategy, and select the optimal compensation strategy to reduce the negative impact brought by data missing;

[0071] After the system completes the acquisition of real-time data such as photovoltaic power generation, load demand, and battery status, and cross-verifies the data consistency through multi-source data fusion technology, the system will use adaptive threshold detection and statistical analysis methods to identify and eliminate abnormal data. However, the elimination of abnormal data may lead to information loss and affect subsequent AI scheduling decisions. Therefore, the system first needs to classify the abnormal data and analyze its impact on the scheduling strategy. For example, if the abnormal data belongs to short-term sensor jitter (such as sudden increase in instantaneous power), it can be directly processed by smoothing filtering; if it is long-term data loss (such as SOC loss caused by communication failure), more complex regression analysis and physical modeling are required for correction. By classifying the abnormal data, the system can select the optimal compensation strategy for different types of data missing situations to minimize the impact of data missing on scheduling decisions.

[0072] Use the regression model to predict and correct the abnormal data to make it conform to the historical trend and improve the accuracy and reliability of the data;

[0073] For abnormal values caused by data missing or errors, historical data can be used for regression analysis and correction. First, the system will search for similar historical data in the context time window of the abnormal data and establish a regression model (such as linear regression, polynomial regression, or deep learning models such as LSTM) to predict the missing data. For example, when correcting the abnormal photovoltaic power generation, the regression model can be trained with photovoltaic data under similar weather conditions in recent days, and multivariate fitting can be carried out in combination with external factors such as temperature and irradiance to calculate a reasonable correction value. At the same time, to enhance the stability of the prediction, the system can use the weighted moving average method (WMA) to weight and combine multiple prediction values, thereby reducing the impact of a single model error on the final data correction. This process ensures that the corrected data matches the historical trend, thereby improving the authenticity and reliability of the data.

[0074] Rationally constrain the regression-corrected data through physical modeling to ensure that the corrected data conforms to physical laws and avoid the impact of incorrect compensation on system operation;

[0075] Although regression analysis can provide reasonable corrected values for abnormal data, relying solely on statistical methods may lead to the corrected data violating physical laws. Therefore, this step introduces physical modeling for constraint correction. For example, in an energy storage system, the change in battery SOC is affected by factors such as charge-discharge efficiency, current, and voltage. Thus, a battery dynamic model (such as Coulomb integration method or Thevenin equivalent circuit model) can be constructed to limit the reasonable range of corrected values. Suppose regression analysis predicts that the SOC should be 60%, but according to physical modeling calculations, the SOC can only reach a maximum of 55% under the current load conditions. Then the corrected value should be adjusted to 55%. In addition, if the predicted value of photovoltaic power generation is higher than the maximum power output allowed by the physical model, the system should be restricted according to the battery charge-discharge curve to avoid overestimating the power generation. This strategy ensures that the corrected data not only conforms to historical trends but also meets physical constraints, avoiding the impact of incorrect compensation on the stable operation of the energy storage system.

[0076] Use linear interpolation, spline interpolation, and Kalman filtering methods to fill in short-term missing data, and optimize compensation in combination with physical constraints to ensure data continuity and scheduling stability;

[0077] For missing data within a short time window, the system uses interpolation compensation technology to reduce the impact of data gaps on the AI scheduling model. First, select an appropriate interpolation method according to the time span of data missing, such as:

[0078] Linear Interpolation: Suitable for situations where data changes smoothly in a short period, such as missing compensation for photovoltaic power generation when there is no sudden weather change.

[0079] Spline Interpolation: Suitable for situations where data changes are more complex, such as compensating the change curve of battery SOC during dynamic charge and discharge.

[0080] Kalman Filter: Suitable for situations where multiple variables change in coordination, such as joint compensation of voltage, current, and SOC.

[0081] Use a time series prediction model (such as LSTM or Transformer) to predict the photovoltaic power generation, load demand, and energy storage status in the future, establish a scheduling warning mechanism, and adjust the energy distribution strategy in advance when sudden abnormalities are predicted to reduce the impact of abnormalities on the stability of the energy storage system;

[0082] Use a time series prediction model (such as LSTM or Transformer) to predict the photovoltaic power generation, load demand, and energy storage state in the future, and establish a scheduling warning mechanism. When sudden anomalies are predicted, adjust the energy allocation strategy in advance to reduce the impact of the anomalies on the stability of the energy storage system. The specific steps are as follows:

[0083] Sort and normalize the data of photovoltaic power generation, load demand, and energy storage state, and use the time window technique to capture long-term and short-term trends to ensure stable data quality;

[0084] After completing the data anomaly detection and correction, the system has obtained high-quality and continuous data on photovoltaic power generation, load demand, and energy storage state. However, these data can only reflect the current and historical operating states and cannot provide a forward-looking decision-making basis for scheduling. Therefore, the goal of this step is to construct an input data set for the time series prediction model so that it can learn historical data patterns and predict the energy state in the future. First, the system will select key feature variables, including photovoltaic power generation, ambient temperature, solar irradiance, battery SOC, grid electricity price, load demand, etc., and normalize the data to eliminate the numerical dimension differences between different variables. Subsequently, the system will construct a time window and use the data of the past period (such as the data of the past 24 hours or 7 days) as the model input to capture long-term and short-term time series features. In addition, for the missing parts after anomaly data correction, the system can use the sliding window filling technique so that the model will not be affected by the missing data at individual time points and affect the overall learning ability.

[0085] Select a time series prediction model for training, optimize the hyperparameters and perform cross-validation to improve the prediction accuracy of photovoltaic power generation, load demand, and energy storage state;

[0086] After constructing the input data set, the system needs to train a time series prediction model to predict future photovoltaic power generation, load demand, and energy storage status. Common time series prediction models include Long Short-Term Memory Network (LSTM), Transformer, ARIMA (Autoregressive Integrated Moving Average Model), etc. LSTM is suitable for cases with long-term time series dependencies and can capture the periodic changes in photovoltaic power generation and load demand; Transformer is better at handling complex time series relationships and is suitable for long-term prediction; ARIMA is suitable for short-term trend prediction with relatively low computational costs. During model training, the system will use historical data for supervised learning and use Mean Squared Error (MSE) or Root Mean Squared Error (RMSE) as the loss function to optimize the model parameters. At the same time, the system can adopt Bayesian optimization methods to adjust hyperparameters (such as learning rate, time window size) to improve the prediction accuracy of the model. In addition, to enhance the generalization ability of the model, the system will perform cross-validation, that is, divide the data set into a training set and a validation set and train at different time periods to ensure that the model can adapt to various operating conditions.

[0087] Set the warning threshold based on the prediction results, dynamically adjust the energy allocation before the decline of photovoltaic power generation or the surge of load, and optimize the scheduling strategy by combining reinforcement learning;

[0088] Based on the trained time series prediction model, the system can predict key parameters such as future photovoltaic power generation, load demand, and battery SOC, thereby establishing a scheduling warning mechanism to cope with sudden abnormal situations. For example, when the system predicts that the photovoltaic power generation is about to drop significantly (such as when rainy weather is approaching) or the load demand surges (such as during the peak period of a communication base station), it can adjust the energy allocation strategy in advance. In specific implementation, the system will set a warning threshold. For example, if it is predicted that the photovoltaic power will drop by more than 30% within the next 1 hour, or the load demand will increase by more than 20% within 30 minutes, a warning will be triggered. At the same time, the system can use reinforcement learning methods combined with the prediction results to dynamically adjust the battery charge and discharge plan, such as charging in advance during low electricity price periods to ensure sufficient power supply during future high-demand periods. In addition, the system can establish a knowledge base based on historical abnormal situations, and further optimize the warning mechanism by comparing the current prediction results with historical abnormal patterns to make it more accurate in predicting possible abnormal conditions.

[0089] Optimize the energy storage scheduling strategy before the occurrence of an anomaly, and adopt a model adaptive adjustment mechanism to correct the prediction error, improving the stability and prediction ability of the energy storage system;

[0090] After predicting a sudden anomaly, the system needs to quickly adjust the scheduling strategy and make real-time corrections when the anomaly occurs. For example, if it is predicted that the photovoltaic power generation will decrease by 50% in the next hour, the system can increase the charging power of the battery in advance before the anomaly occurs, or adjust the power supply strategy on the load side to reduce the dependence on the external power grid. In addition, to ensure that the scheduling strategy is always effective during execution, the system can adopt a model adaptive adjustment mechanism, that is, re-evaluate the prediction error after the scheduling is executed. If the actual situation deviates too much from the predicted value, the parameters of the prediction model are adjusted to improve the accuracy of the model in the future. For example, if the load prediction error continues to be high, the system can adjust the time window length or introduce new feature variables (such as temperature changes) to optimize the prediction model. In addition, after an abnormal event occurs, the system can adopt a self-learning mechanism to incorporate the new abnormal pattern into the training data to improve the future prediction and scheduling ability for similar anomalies.

[0091] Combined with the reinforcement learning algorithm, construct a reward function that includes factors such as battery life, grid electricity price, and load fluctuation, and optimize the energy scheduling strategy based on the distributed reinforcement learning framework to achieve dynamic optimal scheduling of the energy storage system under different working conditions;

[0092] The specific steps to combine the reinforcement learning algorithm, construct a reward function that includes factors such as battery life, grid electricity price, and load fluctuation, and optimize the energy scheduling strategy based on the distributed reinforcement learning framework to achieve dynamic optimal scheduling of the energy storage system under different working conditions are as follows;

[0093] Combined with the time series prediction results, define the state space of reinforcement learning, including key parameters such as photovoltaic power generation, load demand, battery SOC, and grid electricity price, to construct a dynamic scheduling environment;

[0094] After completing the training of the time series prediction model and the establishment of the scheduling early warning mechanism, the system can already predict future photovoltaic power generation, load demand, and energy storage status, and adjust the energy distribution strategy in case of sudden anomalies. However, the scheduling strategy is still based on empirical rules or simple optimization methods and lacks global optimality. Therefore, the goal of this step is to construct a dynamically optimized energy scheduling strategy based on reinforcement learning. First, it is necessary to define the reinforcement learning environment, where the state space (StateSpace) should include key parameters such as the predicted value of photovoltaic power generation, the predicted value of the load, battery SOC, grid electricity price, and historical charge and discharge behavior to comprehensively describe the current state of the energy storage system. At the same time, the state variables need to be standardized to ensure that the features of different numerical magnitudes have an equal impact on the model. In addition, the system can use dimensionality reduction techniques (such as PCA or Autoencoder) to reduce the redundancy of state variables and improve the learning efficiency.

[0095] Design a reward function considering comprehensive economy, battery life, load stability, and renewable energy utilization rate to guide the agent to learn the optimal energy scheduling strategy and dynamically adjust the weights to adapt to different application scenarios;

[0096] Under the reinforcement learning framework, the agent continuously optimizes the energy scheduling strategy through trial and error learning, and the reward function determines the goal of system optimization. Therefore, the core of this step is to design a reward function that balances multiple factors, enabling the energy storage system to achieve dynamic optimal scheduling under different operating conditions. The reward function can include the following key factors:

[0097] 1. Economy: According to the grid electricity price signal, charge preferentially during low electricity price periods and discharge preferentially during high electricity price periods to maximize economic benefits.

[0098] 2. Battery life: Consider the DOD (depth of discharge) and charge-discharge cycles of the battery to avoid overcharging and over-discharging and extend the battery life.

[0099] 3. Stability: Reduce the impact of load fluctuations on the energy storage system, ensure that the SOC is within a reasonable range, and avoid overcharging or over-discharging.

[0100] 4. Renewable energy utilization rate: Improve the local consumption rate of photovoltaic power generation and reduce the phenomenon of curtailment of light.

[0101] Adopt reinforcement learning algorithms such as PPO / DDPG and use the distributed reinforcement learning framework to accelerate training, enabling the agent to learn the optimal energy scheduling strategy under different operating conditions and improving the learning efficiency and convergence speed;

[0102] After constructing the reinforcement learning environment and designing the reward function, the system needs to select an appropriate reinforcement learning algorithm and train it based on a distributed architecture to accelerate policy optimization. Common reinforcement learning algorithms include Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Deep Deterministic Policy Gradient (DDPG). Among them, DQN is suitable for discrete action spaces, while PPO and DDPG are suitable for continuous action spaces. In the energy scheduling problem, the battery charge and discharge power are usually continuous variables, so PPO or DDPG has more advantages. In addition, to improve the training efficiency, the system can adopt a distributed reinforcement learning framework (such as Ape-X or RayRLlib) to train multiple agents in parallel on multiple computing nodes, enabling them to explore the optimal policy simultaneously under different working conditions. For example, one agent can learn the scheduling policy under high light conditions, and another agent can learn the policy under low light conditions at night. Finally, the optimal policies under different working conditions are integrated through federated learning. In addition, the system can also use Experience Replay and Target Network technologies to improve the stability and convergence speed of reinforcement learning.

[0103] Through the online learning and model adaptive adjustment mechanism, the reinforcement learning agent can continuously optimize the scheduling policy, dynamically adapt to environmental changes and sudden anomalies, and improve the stability and interpretability of the energy storage system;

[0104] After the reinforcement learning agent completes training, the system needs to continuously optimize and adjust the scheduling policy during actual operation to adapt to new environmental changes and sudden anomalies. For example, if the system predicts a sharp drop in photovoltaic power generation in the next hour, the scheduling policy needs to adjust the battery charge and discharge plan in advance to ensure stable power supply. In addition, the system can use the Online Learning mechanism to enable the agent to continuously update the policy during actual operation to adapt to dynamic changes such as grid price fluctuations and new device access. For example, when the SOC estimation error increases due to battery aging, the system can retrain the prediction model through the Adaptive Model Adjustment mechanism of reinforcement learning to optimize the energy scheduling. In addition, to enhance the interpretability of the policy, the system can adopt a reinforcement learning method based on the Attention Mechanism, enabling the agent to analyze which state variables have the greatest impact on the decision, thereby improving the transparency and controllability of the policy.

[0105] During the execution of the AI scheduling, by combining confidence estimation and Bayesian optimization methods, dynamically monitor the reliability of the scheduling decision. When the confidence is lower than the set threshold, trigger the rule correction mechanism based on the expert system to reasonably adjust the scheduling result and ensure robustness in case of anomalies;

[0106] During the execution of AI scheduling, by combining confidence estimation and Bayesian optimization methods, the reliability of scheduling decisions is dynamically monitored. When the confidence level is lower than the set threshold, a rule correction mechanism based on an expert system is triggered to reasonably adjust the scheduling results and ensure robustness in abnormal situations. The specific steps are as follows:

[0107] During the execution of Al scheduling, first evaluate the confidence level of the current scheduling strategy to judge the reliability of the decision. Confidence estimation uses Bayesian neural network (BNN) or Monte Carlo (MonteCarloDropout) methods. By performing multiple forward propagations to calculate the mean and variance of the scheduling strategy, the uncertainty of the scheduling strategy is quantified. The formula is as follows:

[0108]

[0109] In the formula, C t is the confidence level of the scheduling decision, with a value range of [0, 1]. The larger the value, the higher the decision reliability. σ[P t is the uncertainty (standard deviation) of the current scheduling decision, calculated by the Bayesian neural network. E[P t is the expected value of the scheduling decision at the current moment, that is, the average decision value in multiple forward propagations. ∈ is a very small positive number to prevent the denominator from approaching zero;

[0110] When the confidence level C t is lower than the set threshold C thresh (e.g., C t <0.7), then the rule correction mechanism is triggered to ensure the stability of the scheduling strategy;

[0111] Calculate the confidence level C t of the scheduling decision. If it is lower than the set threshold C thresh , then the scheduling decision needs to be further optimized to enhance the system robustness.

[0112] When the confidence level is lower than the threshold C thresh , optimize the current scheduling parameters (such as the battery charge and discharge power P t ) to find the optimal scheduling value. Use Bayesian optimization (Bayesian Optimization, BO) to dynamically adjust the scheduling strategy, so that the optimized decision reduces risks while improving economy and energy storage system stability. The objective function of Bayesian optimization is defined as follows:

[0113]

[0114] In the formula, $P_{opt}$ is the optimized battery charge and discharge power, $R(P)$ is the reward function considering factors such as grid electricity price, battery life, and load fluctuations, and $\lambda$ is the risk weight coefficient that controls the degree of penalty for uncertainty. $P_{opt}$ is the value of $P$ that maximizes the objective function.

[0115] Bayesian optimization constructs a surrogate model through Gaussian Process Regression (GPR) and uses the Expected Improvement (EI) or Probability of Improvement (PI) criterion to select the optimal scheduling parameters. After optimization, the system will use $P_{opt}$ as the corrected scheduling value and input it into the next rule correction mechanism. Summary: The Bayesian optimization method is used to adjust the scheduling parameters. While optimizing the reward function, it reduces the scheduling uncertainty and improves the reliability and economy of the scheduling scheme.

[0116] Although Bayesian optimization provides optimized scheduling parameters, it may still lead to unacceptable scheduling decisions in extreme cases (such as sudden load surges or the battery SOC exceeding the safe range). Therefore, a rule correction mechanism based on an expert system is adopted to perform final correction to ensure the safety and feasibility of the scheduling strategy. The expert system is based on a set of predefined rules. For example, if the battery SOC is below 20% and it is still discharging, then $P$ is forced to be set to 0; if the grid electricity price is much lower than the historical average, then increasing the charging power is encouraged; if the load prediction value fluctuates violently in a short period of time, then $P$ is smoothed. The rule correction formula is expressed as follows: In the formula,

[0117]

[0118] where $P$ is the finally executed battery charge and discharge power, $I(·)$ is the indicator function that takes 1 if the condition holds and 0 otherwise, $\alpha$ is the smoothing weight factor, and $\Delta P$ smooth is the smoothing adjustment value based on load fluctuations.

[0119] Finally, the system adopts $P$ as the new scheduling strategy and feeds it back to the time series prediction module to update the calculation basis for future scheduling decisions.

[0120] Adopt a federated learning architecture to enable multiple base stations to share abnormal data patterns and collaboratively optimize the AI model while protecting data privacy; through an active learning mechanism, dynamically expand the training set, enhance the model's adaptability to data noise and outliers, and improve the robustness of the overall scheduling system;

[0121] The specific steps to adopt a federated learning architecture to enable multiple base stations to share abnormal data patterns and collaboratively optimize the AI model while protecting data privacy; through an active learning mechanism, dynamically expand the training set, enhance the model's adaptability to data noise and outliers, and improve the robustness of the overall scheduling system are as follows:

[0122] During the optimization and scheduling process of the hybrid energy storage system, the data of a single base station is limited, and the abnormal data patterns vary due to various factors such as geographical location, climate, and load characteristics. Therefore, to improve the generalization ability of the Al scheduling model, a federated learning (FederatedLearningFL) architecture is adopted to enable multiple base stations to share abnormal data patterns while ensuring that data privacy is not leaked. Specifically, each base station maintains a local Al model, and after one round of local training, it sends the model gradient rather than the original data to the central server. The server aggregates the model gradients of multiple base stations based on the federated average (FedAvg) algorithm and updates the global model parameter θ. The core calculation formula for federated model aggregation is as follows:

[0123]

[0124] In the formula, θ (j+1) is the global model parameter in the (j + 1)-th round, N is the total number of base stations, and each base station independently trains its local model. w i is the data volume of base station i, representing the number of data samples collected by this base station. W is the normalization factor of the total data volume of all base stations. is the model parameter of base station i during the j-th round of training. γ is the weight coefficient of abnormal data for model update (used to enhance the model's adaptability to abnormal data), and M is the normalization factor of the total abnormal data volume of all base stations. is the model gradient of base station i after local training;

[0125] This step optimizes the global model through a dual-weighting mechanism (based on data volume and the proportion of abnormal data), making base stations with more abnormal data contribute more to the global model, enhancing the model's adaptability to abnormal data patterns, and at the same time preventing the models of small data sets from having too much impact on global optimization.

[0126] Since the adaptability of the energy storage system scheduling AI model to abnormal data depends on the diversity and coverage of the training data, after optimizing the global model through federated learning, an Active Learning mechanism is further adopted to dynamically expand the training set and improve the robustness of the model to noise and outliers. Specifically, after each round of training, the system calculates the uncertainty of the current model and selects the data samples with the largest amount of information to add to the training set, thereby improving the generalization ability of the model in abnormal scenarios. The sample selection for active learning is based on the principle of Maximum Information Gain (MIG), and the calculation formula is as follows:

[0127]

[0128] In the formula, S * is the subset of active learning samples selected, S c is the candidate data set, which contains all samples to be screened, x represents a certain data point in the candidate sample set S c , p(y|x, θ (j+1) ) is the predicted probability distribution of sample x under the global model parameters θ (j+1) , representing the uncertainty of this sample, H(p(y|x, θ (j+1) )) is the prediction entropy of the current global model, K is the number of sub-models used to calculate the uncertainty, represents the parameters of the k-th sub-model after the (j + 1)-th round of training, is to maximize the information gain and select valuable samples.

[0129] The core idea of the above steps is to calculate the change in the prediction entropy of the samples, that is, to select the samples that cause the largest decrease in the entropy of the global model to add to the training set to maximize the information gain of the model. The newly selected samples S * will be fed back to each base station for local training, and the global model will be optimized again through the federated learning framework to form a dynamic loop optimization mechanism.

[0130] Embodiment 1: In a photovoltaic energy storage base station, the accuracy and continuity of data are crucial for intelligent scheduling. Since the energy storage system involves multiple data sources, including photovoltaic power generation, battery SOC (state of charge), load demand, etc., data anomalies may occur during the data acquisition process due to factors such as sensor failures, communication delays, and environmental interference. To ensure the stability of the scheduling system, this embodiment first uses multi-source data fusion and adaptive threshold detection techniques to monitor the collected data in real time and eliminate anomalies. For example, the system can compare the data consistency of multiple sensors. If there is a significant deviation between the power output of the photovoltaic MPPT controller and the power measured by the electricity meter, it may indicate that a certain sensor has failed. In addition, the system can also use statistical methods such as moving window mean and Z-score analysis to identify anomaly points, and combine time series prediction to detect short-term mutations to improve data quality.

[0131] After data cleaning, the system needs to correct and compensate for the abnormal data. For the marked abnormal data, a fusion algorithm combining historical data regression analysis and physical modeling can be used for correction. For example, in the case of missing SOC data, the system can predict a reasonable SOC value based on the SOC change trend under similar load conditions in recent days using a regression model (such as polynomial regression or LSTM). At the same time, physical constraint correction is performed in combination with the battery charge and discharge model (such as the Thevenin equivalent circuit model) to ensure that the corrected SOC value does not exceed the physically possible range. In addition, for missing data within a short time window, interpolation compensation methods such as linear interpolation or spline interpolation can be used to keep the data smooth and reduce the impact on scheduling decisions.

[0132] After ensuring data quality, the system needs to predict photovoltaic power generation, load demand, and battery SOC for a period of time in the future to provide forward-looking guidance for scheduling decisions. The system can use time series prediction models such as LSTM (long short-term memory network), Transformer, or ARIMA (autoregressive integrated moving average model). Among them, LSTM is suitable for capturing load change patterns with long-term dependencies, while Transformer can handle complex time series data and improve prediction accuracy. During system training, historical data can be used for supervised learning, and Bayesian optimization can be used to adjust hyperparameters to improve the generalization ability of the model. Based on the prediction results, the system establishes a scheduling warning mechanism to adjust the battery charge and discharge plan in advance when photovoltaic power generation drops sharply or the load surges to ensure the stable operation of the energy storage system.

[0133] The advantage of this embodiment is that it not only improves the reliability of data, but also adjusts the scheduling strategy in advance through time series prediction, reducing the impact of anomalies on the stability of the energy storage system. This method is applicable to scenarios with unstable photovoltaic power generation and large fluctuations in load demand, and can effectively improve the adaptability and operating efficiency of the energy storage system.

[0134] Embodiment 2: After completing data cleaning and time series prediction, the system already has high-quality input data and is able to predict future photovoltaic power generation, load demand, and battery SOC. However, traditional rule-based scheduling strategies usually rely on manual experience and are difficult to adapt to complex dynamic environments. Therefore, this embodiment uses a reinforcement learning (RL) algorithm to optimize the energy scheduling strategy of the energy storage system.

[0135] First, it is necessary to construct a reinforcement learning environment and define the state space, action space, and reward function. The state space includes key parameters such as predicted photovoltaic power generation value, predicted load value, battery SOC, grid electricity price, and historical charge and discharge behavior to comprehensively describe the current state of the energy storage system. The action space defines the executable scheduling strategies, such as charging power, discharging power, and grid power purchase ratio. The reward function is the basis for the agent to optimize the strategy and usually includes the following key factors:

[0136] Economy: Charge preferentially during low electricity price periods and discharge preferentially during high electricity price periods to maximize economic benefits.

[0137] Battery life: Consider the depth of discharge (DOD) and the number of cycles, and avoid overcharging and over-discharging to extend the battery life.

[0138] System stability: Reduce the impact of load fluctuations on the energy storage system, ensure that the SOC is maintained within a reasonable range, and prevent overcharging or over-discharging.

[0139] Renewable energy utilization rate: Improve the local consumption rate of photovoltaic power generation and reduce the phenomenon of curtailment of light.

[0140] In terms of algorithm selection, the system can adopt reinforcement learning methods such as PPO (Proximal Policy Optimization) or DDPG (Deep Deterministic Policy Gradient). Among them, PPO is suitable for high-dimensional continuous action spaces, while DDPG is suitable for low-latency real-time control tasks. In addition, to accelerate training, the system can adopt a distributed reinforcement learning framework (such as Ape-X or Ray RLlib) to parallel train the agent on multiple computing nodes, enabling it to explore the optimal strategy simultaneously under different operating conditions. In addition, combined with an online learning mechanism, the system can continuously optimize the scheduling strategy during actual operation to dynamically adapt to environmental changes and improve the level of intelligence.

[0141] The advantage of this embodiment is that it can automatically learn the optimal scheduling strategy without manual intervention, is applicable to complex and variable photovoltaic energy storage scenarios, and can significantly improve the economy and stability of the energy storage system.

[0142] Embodiment 3: During the operation of the energy storage system, in addition to daily load fluctuations and changes in photovoltaic power generation, sudden abnormal situations such as extreme weather and battery failures may also occur. To cope with these situations, this embodiment introduces an intelligent abnormal scenario detection and correction mechanism to improve the adaptability and robustness of the system.

[0143] First, the system uses confidence estimation and Bayesian optimization methods to evaluate the reliability of the scheduling strategy in real time. For example, during the execution of AI scheduling, if the system detects that the current load prediction value deviates too much from the actual load (exceeding twice the standard deviation of the historical average error), it indicates that the prediction model may fail and an abnormal correction mechanism needs to be triggered. In addition, if the confidence level of the photovoltaic power generation prediction result is low (such as a large variance in LSTM prediction), the system can reduce its reliance on the prediction data and adopt a more conservative scheduling strategy.

[0144] After triggering the abnormal correction mechanism, the system can adopt an expert system rule correction mechanism to optimize the scheduling strategy based on historical data and operation and maintenance experience. For example, when the photovoltaic power supply is insufficient due to long-term rainy weather, the system can increase the proportion of power purchased from the grid in advance and reduce the power supply priority of non-essential loads. In addition, the system can combine the federated learning architecture to enable multiple base stations to share abnormal data patterns and collaboratively optimize the AI model without leaking data privacy, improving the overall system's ability to handle abnormal situations. For example, if a base station has experienced extreme high-temperature weather, its abnormal handling experience can be used to optimize the scheduling strategies of other base stations and improve the intelligent scheduling ability of the entire network.

[0145] The advantage of this embodiment is that it can effectively cope with sudden abnormal situations, avoid system instability caused by prediction errors or model failures, ensure the long-term reliable operation of the photovoltaic energy storage system, and is applicable to high-demand and high-reliability energy management scenarios.

[0146] Through data cleaning, time series prediction, and reinforcement learning to optimize the scheduling strategy, the energy storage system can achieve intelligent management in complex environments and enhance its adaptability to dynamic changes. First, multi-source data fusion and anomaly detection technologies are adopted to ensure the accuracy of key data such as photovoltaic power generation, load demand, and battery SOC, avoiding problems such as sensor failures and communication delays from affecting scheduling decisions. Second, with the help of time series prediction models such as LSTM and Transformer, the system can predict future photovoltaic power generation and load demand in advance and establish a scheduling warning mechanism to optimize the energy storage strategy before anomalies occur, preventing power supply instability caused by sudden load shocks. In addition, by combining reinforcement learning to optimize the scheduling strategy, the system can autonomously learn the optimal scheduling plan based on grid electricity prices, photovoltaic output, and load fluctuations. For example, it reduces discharging during peak electricity prices to reduce battery loss while ensuring stable power supply. Through intelligent optimization, the system can dynamically adjust the strategy under different climate conditions, load patterns, and energy storage states, enhancing the self-adaptability of the energy storage system and improving the long-term stable operation ability of the photovoltaic energy storage base station.

[0147] Through intelligent regulation, the system can comprehensively consider factors such as grid electricity prices, photovoltaic power generation prediction, and battery life, and dynamically adjust the charge and discharge plan, thereby reducing operating costs. For example, the system can preferentially charge during low electricity price periods and release energy storage during peak electricity prices, reducing grid power purchase expenditures while increasing the local consumption rate of photovoltaic power generation. In addition, to prevent the battery from aging due to frequent deep charge and discharge, the system optimizes the scheduling through reinforcement learning, intelligently controls the battery SOC within the optimal working range of 40%-80%, reduces overcharging and over-discharging, and improves the battery health (SOH). At the same time, combined with the abnormal scenario detection and intelligent correction mechanism, when detecting extreme load fluctuations, battery aging, or abnormal environmental temperature, the system can automatically adjust the scheduling strategy to prevent the battery from overheating or discharging abnormally, improving battery life and reducing long-term operation and maintenance costs.

[0148] All the above formulas are dimensionless and take their numerical values for calculation. The formula is obtained by collecting a large amount of data for software simulation to get a formula closest to the actual situation. The preset parameters in the formula are set by those skilled in the art according to the actual situation.

[0149] Only some exemplary embodiments of the present invention have been described by way of illustration. Undoubtedly, for those of ordinary skill in the art, without departing from the spirit and scope of the present invention, the described embodiments can be modified in various different ways. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0150] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0151] It should be understood that in various embodiments of the present application, the sequence numbers of the above processes do not imply the order of execution, and the order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0152] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. A professional technician can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0153] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0154] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0155] In addition, in each embodiment of the present application, the functional units can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0156] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claimed rights.

[0157] Only some exemplary embodiments of the present invention have been described above by way of illustration. Without doubt, for those of ordinary skill in the art, the described embodiments can be modified in various different ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the protection scope of the claims of the present invention.

Claims

1. The hybrid energy storage system optimization scheduling method based on AI intelligent control is characterized by: The following steps are involved: Collect real-time data and cross-validate data consistency through multi-source data fusion technology; use adaptive threshold detection combined with statistical analysis methods to identify data anomalies caused by sensor failure, communication delays and environmental interference, and eliminate abnormal data; A fusion algorithm based on historical data regression analysis and physical modeling is used to correct the marked abnormal data, and interpolation compensation is performed using adjacent time window data to reduce the impact of data missing on scheduling decisions; Use the time series prediction model to predict the photovoltaic power generation, load demand and energy storage status in the future, establish a dispatch warning mechanism, and adjust the energy allocation strategy in advance when sudden abnormalities are predicted to reduce the impact of abnormalities on the stability of the energy storage system; Combined with reinforcement learning algorithm, a reward function is constructed, and the energy scheduling strategy is optimized based on the distributed reinforcement learning framework, so that the energy storage system can achieve dynamic optimal scheduling under different working conditions; During the AI ​​scheduling execution process, the reliability of scheduling decisions is dynamically monitored by combining confidence estimation with Bayesian optimization methods. When the confidence is lower than the set threshold, the rule correction mechanism based on the expert system is triggered to reasonably adjust the scheduling results to ensure robustness under abnormal circumstances. A federated learning architecture is adopted to enable multiple base stations to share abnormal data patterns and collaboratively optimize AI models while protecting data privacy. The training set is dynamically expanded through an active learning mechanism to improve the model's adaptability to data noise and outliers, thereby improving the robustness of the overall scheduling system.

2. The hybrid energy storage system optimization scheduling method based on AI intelligent control according to claim 1 is characterized in that: Collect real-time data and cross-verify data consistency through multi-source data fusion technology; use adaptive threshold detection combined with statistical analysis methods to identify data anomalies caused by sensor failure, communication delays and environmental interference, and remove abnormal data. The specific steps are as follows: Collect key data of photovoltaic, battery and load, and perform time alignment, format standardization and interpolation processing to ensure data consistency and availability; Adopt multi-source data fusion technology to cross-validate data from different sensors to improve data accuracy; Combine adaptive threshold detection, statistical analysis and time series prediction methods to dynamically identify data anomalies and improve the sensitivity and adaptability of anomaly detection; Abnormal data is filled through sliding window filtering, regression analysis and deep learning prediction, and the compensation strategy is optimized in combination with physical constraints to ensure the stable operation of the energy storage system.

3. The hybrid energy storage system optimization scheduling method based on AI intelligent control according to claim 1 is characterized in that: The fusion algorithm based on historical data regression analysis and physical modeling is used to correct the marked abnormal data, and interpolation compensation is performed using adjacent time window data to reduce the impact of data missing on scheduling decisions. The specific steps are as follows: Classify abnormal data, analyze its impact on scheduling strategies, and select the optimal compensation strategy to reduce the negative impact of missing data; Use regression models to predict and correct abnormal data to make it consistent with historical trends and improve data accuracy and reliability; Physical modeling is used to constrain the rationality of regression correction data to ensure that the corrected data conforms to physical laws and avoids erroneous compensation affecting system operation. Linear interpolation, spline interpolation and Kalman filtering methods are used to fill short-term missing data, and combined with physical constraint optimization compensation to ensure data continuity and scheduling stability.

4. The hybrid energy storage system optimization scheduling method based on AI intelligent control according to claim 1 is characterized in that: The time series prediction model is used to predict the photovoltaic power generation, load demand and energy storage status in the future, and a dispatch warning mechanism is established. When an abnormality is predicted, the energy allocation strategy is adjusted in advance to reduce the impact of the abnormality on the stability of the energy storage system. The specific steps are as follows: Organize and normalize PV power generation, load demand, and energy storage status data, and use time window technology to capture long-term and short-term trends to ensure stable data quality; Select a time series prediction model for training, optimize hyperparameters and perform cross-validation to improve the prediction accuracy of photovoltaic power generation, load demand and energy storage status; Set warning thresholds based on prediction results, dynamically adjust energy allocation before photovoltaic power generation drops or load surges, and optimize scheduling strategies in combination with reinforcement learning; The energy storage scheduling strategy is optimized before an anomaly occurs, and the model adaptive adjustment mechanism is used to correct the prediction error, thereby improving the stability and prediction ability of the energy storage system.

5. The hybrid energy storage system optimization scheduling method based on AI intelligent control according to claim 1 is characterized in that: Combined with the reinforcement learning algorithm, the reward function is constructed, and the energy scheduling strategy is optimized based on the distributed reinforcement learning framework, so that the energy storage system can achieve dynamic optimal scheduling under different working conditions. The specific steps are as follows; Combined with the time series prediction results, the state space of reinforcement learning is defined to build a dynamic scheduling environment; The reward function is designed based on economic efficiency, battery life, load stability and renewable energy utilization, guiding the agent to learn the optimal energy scheduling strategy and dynamically adjust the weights to adapt to different application scenarios; Adopt reinforcement learning algorithms such as PPO / DDPG, and use the distributed reinforcement learning framework to accelerate training, so that the intelligent agent can learn the optimal energy scheduling strategy under different working conditions, improving learning efficiency and convergence speed; Through online learning and model adaptive adjustment mechanisms, the reinforcement learning agent can continuously optimize the scheduling strategy and dynamically adapt to environmental changes and sudden anomalies, thereby improving the stability and explainability of the energy storage system.

6. The hybrid energy storage system optimization scheduling method based on AI intelligent control according to claim 1 is characterized in that: During the AI ​​scheduling execution process, the reliability of scheduling decisions is dynamically monitored by combining confidence estimation with the Bayesian optimization method. When the confidence is lower than the set threshold, the rule correction mechanism based on the expert system is triggered to reasonably adjust the scheduling results. The specific steps to ensure robustness under abnormal conditions are as follows: During the execution of Al scheduling, the confidence of the current scheduling strategy is first evaluated to determine the reliability of the decision. The mean and variance of the scheduling strategy are calculated through multiple forward propagations to quantify the uncertainty of the scheduling strategy. The formula is as follows: In the formula, C t is the confidence of the scheduling decision, σ[P t ] is the uncertainty of the current scheduling decision, E[P t ] is the expected value of the scheduling decision at the current moment, ∈ is a very small positive number to prevent the denominator from approaching zero; When the confidence level C t Below the set threshold C thresh When , the rule correction mechanism is triggered to ensure the stability of the scheduling strategy; When the confidence level is lower than the threshold C thresh When the current dispatching parameters are optimized to find the optimal dispatching value, the dispatching strategy is dynamically adjusted using Bayesian optimization, so that the optimized decision reduces the risk and improves the economy and stability of the energy storage system. The objective function of Bayesian optimization is defined as follows: In the formula, is the optimized battery charging and discharging power, R(P) is the reward function, which takes into account factors such as grid electricity price, battery life, load fluctuation, etc., λ is the risk weight coefficient, which controls the degree of penalty for uncertainty. is the value of P that maximizes the objective function.

7. The hybrid energy storage system optimization scheduling method based on AI intelligent control according to claim 6 is characterized in that: Although Bayesian optimization provides optimized scheduling parameters However, it may still lead to unacceptable scheduling decisions in extreme cases. Therefore, a rule correction mechanism based on an expert system is adopted to correct Perform final correction to ensure the safety and feasibility of the scheduling strategy. The rule correction formula is expressed as follows: In the formula, is the final battery charge and discharge power, I(·) is the indicator function, which takes 1 if the condition is met, otherwise it takes 0, α is the smoothing weight factor, ΔP smooth It is a smooth adjustment value based on load fluctuation.

8. The hybrid energy storage system optimization scheduling method based on AI intelligent control according to claim 1 is characterized in that: The federated learning architecture is adopted to enable multiple base stations to share abnormal data patterns and collaboratively optimize the AI ​​model while protecting data privacy. The specific steps to dynamically expand the training set through the active learning mechanism to improve the model's adaptability to data noise and outliers and improve the robustness of the overall scheduling system are as follows: In the process of optimizing the scheduling of hybrid energy storage systems, the data of a single base station is limited, and the abnormal data patterns vary due to factors such as geographical location, climate, and load characteristics. Therefore, in order to improve the generalization ability of the Al scheduling model, a federated learning architecture is adopted to enable multiple base stations to share abnormal data patterns while ensuring that data privacy is not leaked. The core calculation formula for federated model aggregation is as follows: In the formula, θ (j+1) is the global model parameter of the j+1th round, N is the total number of base stations, and w i is the amount of data of base station i, W is the normalization factor of the total amount of data of all base stations, is the model parameter of base station i in the jth round of training, γ is the weight coefficient of abnormal data on model update, M is the normalization factor of the total amount of abnormal data of all base stations, is the model gradient of base station i after local training; Since the adaptability of the energy storage system scheduling AI model to abnormal data depends on the diversity and coverage of the training data, after optimizing the global model through federated learning, an active learning mechanism is further adopted to dynamically expand the training set and improve the robustness of the model to noise and outliers. The sample selection of active learning is based on the principle of maximizing information gain, and the calculation formula is as follows: In the formula, S * is the selected active learning sample subset, S c is a candidate data set, containing all samples to be screened, and x represents the candidate sample set S c A data point in , p(y|x, θ (j+1) ) is the global model parameter θ (j+1) Under this condition, the predicted probability distribution of sample x is H(p(y|x,θ (j+1) )) is the prediction entropy of the current global model, K is the number of sub-models used to calculate uncertainty, represents the parameters of the k-th sub-model after j+1 rounds of training, The purpose is to maximize information gain and select valuable samples.

Images