Energy management system of light storage and charging integrated device

Abstract

The invention relates to the technical field of new energy, and discloses an optical storage and charging integrated device energy management system, which comprises a data acquisition module used for acquiring a multi-source heterogeneous original data stream and carrying out preprocessing and state feature extraction; the prediction module is used for intelligently predicting photovoltaic output and charging load; the optimization scheduling module is used for carrying out multi-target energy optimization scheduling; the real-time control module is used for carrying out equipment power control and coordination operation; the health management module is used for performing energy storage health state evaluation and system performance analysis operation; the strategy learning module is used for acquiring operation scene features and carrying out control strategy learning optimization and operation effect verification; according to the invention, by constructing the multi-module collaborative intelligent energy management system, the photovoltaic consumption rate is improved, the operation cost is reduced, the adaptability and reliability are enhanced, and powerful technical support is provided for large-scale application of the photovoltaic storage and charging integrated energy station.

Description

Technical Field

[0001] This invention relates to the field of new energy technology, and more specifically, to an energy management system for an integrated photovoltaic, energy storage, and charging device. Background Technology

[0002] With the rapid growth of electric vehicle ownership and the continuous expansion of renewable energy power generation, integrated photovoltaic-storage-charging energy stations have received widespread attention and application as a new type of energy infrastructure. These systems integrate photovoltaic power generation, energy storage systems, and charging pile facilities, enabling the local consumption and efficient utilization of clean energy. However, existing integrated photovoltaic-storage-charging systems suffer from numerous technical deficiencies in energy management, severely restricting the overall performance and economic benefits of the system.

[0003] The main problem with existing technologies lies in the lack of effective multi-source heterogeneous data fusion and intelligent prediction mechanisms. Traditional systems often employ simple data acquisition methods, which cannot effectively process multi-source heterogeneous data from different devices such as photovoltaic arrays, energy storage systems, and charging pile clusters, leading to inaccurate system status perception. Simultaneously, existing prediction methods largely rely on single models, resulting in low accuracy in predicting photovoltaic output and charging load, especially under complex weather conditions and special scenarios such as holidays, where prediction errors are significant and cannot provide reliable decision-making basis for energy dispatch. Furthermore, traditional systems lack accurate assessment and adaptive management capabilities for the health status of energy storage. Energy storage systems often operate according to fixed parameters, failing to dynamically adjust operating strategies based on actual health status, leading to shortened energy storage lifespan and reduced system regulation capabilities.

[0004] More importantly, existing systems generally lack autonomous learning and strategy optimization capabilities. Control strategies are mostly based on manually set fixed parameters, unable to adaptively adjust according to operational experience and environmental changes. This static control mode often fails to achieve optimal energy allocation and equipment coordination in complex and ever-changing real-world operating scenarios, leading to problems such as low photovoltaic absorption rates, high operating costs, and poor equipment utilization efficiency. Especially in application scenarios with large load fluctuations and limited grid connection capacity, such as highway service areas, traditional energy management systems struggle to balance multiple objectives such as economic efficiency, environmental friendliness, and equipment lifespan, requiring improvement in overall system performance. Summary of the Invention

[0005] This invention provides an energy management system for an integrated photovoltaic, energy storage, and charging device, which solves the technical problems of insufficient coordinated control and low energy utilization efficiency of photovoltaic power generation, energy storage system, and charging facilities in related technologies.

[0006] This invention provides an energy management system for an integrated photovoltaic, energy storage, and charging device, comprising: The data acquisition module is used to acquire multi-source heterogeneous raw data streams, perform preprocessing and state feature extraction, and obtain standard operating state feature vectors. The prediction module, based on the feature vector of standard operating state, performs intelligent prediction of photovoltaic output and charging load, and obtains intelligent prediction results; The optimization scheduling module performs multi-objective energy optimization scheduling based on intelligent prediction results and generates real-time scheduling instructions; The real-time control module performs equipment power control and coordination operations based on real-time scheduling instructions, and obtains equipment power control results. The health management module performs energy storage health status assessment and system performance analysis based on equipment power control results and standard operating status feature vectors, and obtains energy storage health status assessment results and health management results. The strategy learning module builds a case library based on health management results, acquires the characteristics of operating scenarios, learns and optimizes control strategies, verifies the operating effects, and obtains intelligent control strategies that evolve autonomously through learning.

[0007] In a preferred embodiment, the data acquisition module includes: Deploy a distributed sensor network, using industrial Ethernet and ModbusTCP communication protocols to synchronously collect electrical and environmental parameters of measurement points at a preset sampling period, and obtain multi-source heterogeneous raw data streams; Time alignment, data normalization, and wavelet transform denoising are performed on the multi-source heterogeneous raw data stream to obtain the denoised data sequence; Based on the denoised data sequence, the extended Kalman filter algorithm is used to estimate the state of charge and obtain the corrected energy storage SOC value. Based on the denoised data sequence and the corrected energy storage SOC value, principal component analysis is used to extract state features and obtain the standard operating state feature vector.

[0008] In a preferred embodiment, the prediction module includes: A photovoltaic power output prediction model is constructed. Historical operating data of the photovoltaic-storage-charging system is collected as the source domain dataset. A long short-term memory network is used to construct the basic photovoltaic power output prediction model. Data collected from the local photovoltaic-storage-charging system is used as the target domain data to obtain the photovoltaic power output prediction model. A multi-step rolling prediction method is used to generate photovoltaic power prediction values, resulting in photovoltaic output prediction curves at multiple time scales.

[0009] In a preferred embodiment, the prediction module further includes: The quantile regression method is used to estimate the prediction uncertainty. In the output layer of the basic photovoltaic power output prediction model, the quantile loss function is used instead of the traditional mean square error loss function. Multiple prediction models with preset quantiles are trained and correspond to the lower limit, median and upper limit of prediction, respectively, to obtain the photovoltaic power output prediction interval. Based on historical charging records, statistical analysis was conducted. A vehicle arrival model was established based on the Poisson process assumption. The probability distribution of charging power and charging time was fitted using the Gaussian kernel density estimation method. The load was aggregated using the Monte Carlo random sampling method to obtain the charging load probability density distribution.

[0010] In a preferred embodiment, the optimized scheduling module includes: A multi-objective optimization model is constructed, which includes economic objectives, environmental objectives, and equipment life objectives. The economic objective function is to minimize the system's daily operating cost, the environmental objective function is to maximize the photovoltaic absorption rate, and the equipment life objective function is to minimize the equivalent cycle number of the energy storage system. The economic objective function, environmental objective function, and equipment lifespan objective function are transformed to a unified dimension by a normalization method. An adaptive weighting coefficient is introduced to construct a comprehensive objective function in the form of a weighted sum. The weighting coefficient is dynamically adjusted according to the current energy storage health status and electricity price level to obtain a single-objective optimization problem.

[0011] In a preferred embodiment, the optimized scheduling module further includes: Establish a set of constraints for day-ahead energy dispatch, including power balance constraints, energy storage SOC constraints, energy storage charging and discharging power constraints, grid interaction power constraints, charging pile output power constraints, and energy storage charging and discharging state mutual exclusion constraints, to obtain a complete day-ahead optimal dispatch mathematical model; The optimal scheduling plan for energy storage charging and discharging power, grid interaction power, and photovoltaic utilization power in each time period of the future preset time period is obtained by using a mixed integer linear programming algorithm. A rolling optimization strategy is adopted to establish a real-time energy scheduling model. A preset time period is used as a time period, and the optimization is performed once every preset time interval to obtain a refined power allocation instruction. The power allocation instruction of the current time period is extracted to obtain the real-time scheduling instruction to be sent to the underlying controller.

[0012] In a preferred embodiment, the real-time control module includes: The power command is sent to the photovoltaic inverter, energy storage converter and charging pile controller using a standardized industrial communication protocol, and feedback signals confirming receipt are obtained from each device. A power regulation algorithm based on droop control is used to control the charging and discharging behavior of the energy storage system. The energy storage converter adopts a dual closed-loop control structure, with the outer loop being the power control loop and the inner loop being the current control loop. When multiple energy storage converters are connected in parallel, the droop control strategy is used to achieve automatic power distribution and obtain the DC bus voltage and actual charging and discharging power.

[0013] In a preferred embodiment, the health management module includes: The charge-discharge cycles of energy storage SOC are counted using the rainflow counting method. The peak and valley values ​​in the SOC sequence are identified. Peak-valley pairs are matched to form cycles according to the rainflow rule. Each cycle is characterized by cycle depth and average SOC, and the cycle number distribution is obtained. The equivalent number of full cycles of the energy storage system is calculated using the linear cumulative damage theory. Based on the cycle life corresponding to different cycle depths, the lifetime damage caused by each actual cycle is equivalent to the number of full cycles, thus obtaining an estimated value of energy storage capacity decay. A comprehensive scoring method is used to calculate the energy storage health score. Based on the score, the energy storage health status is divided into four levels: excellent, good, average, and declining, thus obtaining the energy storage health status index.

[0014] In a preferred embodiment, the health management module further includes: An isolated forest anomaly detection algorithm is used to identify abnormal operating modes of energy storage systems. Multiple isolated trees are constructed to calculate the average path length of data points. When the anomaly score of real-time monitoring data exceeds a preset threshold, an anomaly warning is triggered to obtain a potential fault warning signal. Based on the energy storage health status indicators, an adaptive strategy is used to dynamically correct the operating constraint parameters of the energy storage system. Different SOC operating ranges and charging / discharging power limits are set for energy storage units with different health levels to obtain the energy storage operation strategy.

[0015] In a preferred embodiment, the policy learning module includes: Each system run is recorded as a case, and each case includes the characteristics of the running scenario, control strategy parameters, and running effect indicators, thus obtaining a running case library; Define a similarity measurement function for scene features, calculate the comprehensive similarity between the current scene and historical cases, and select the top few cases with the highest similarity as a set of reference cases; The energy management problem is modeled as a Markov decision process, and a deep Q-network algorithm is used for policy learning. State transitions and rewards are stored in an experience replay pool, and the neural network parameters are updated by randomly sampling training samples to obtain an intelligent control strategy that learns and evolves autonomously.

[0016] The beneficial effects of this invention are as follows: This invention effectively solves key problems in existing technologies, such as insufficient data fusion, low prediction accuracy, and inadequate energy storage management, by constructing a multi-module collaborative intelligent energy management system. The data acquisition module employs multi-source heterogeneous data fusion technology, combined with wavelet transform denoising and principal component analysis feature extraction, to improve the accuracy and real-time performance of system state perception. The prediction module achieves high-precision prediction of photovoltaic output and charging load through transfer learning and multi-model integration methods, providing a reliable data foundation for optimized scheduling. The optimized scheduling module establishes a multi-objective optimization model that can achieve a dynamic balance between economy, environmental protection, and equipment lifespan, ensuring the system maintains its optimal state under various operating scenarios through real-time rolling optimization.

[0017] The health management module and strategy learning module further enhance the system's intelligence and long-term operational performance. The health management module, through rainflow counting and comprehensive health scoring, achieves precise health status assessment and adaptive operation strategy adjustment of the energy storage system, effectively extending the energy storage's lifespan and ensuring system safety. The strategy learning module, based on a case study library and reinforcement learning algorithms, enables the system to autonomously learn and optimize strategies, continuously improving control strategies based on historical operating experience and achieving continuous performance enhancement. Through the coordinated operation of the modules, the entire system improves photovoltaic absorption rate, reduces operating costs, and enhances system adaptability and reliability, providing strong technical support for the large-scale application of integrated photovoltaic-storage-charging energy stations. Attached Figure Description

[0018] Figure 1 This is a block diagram of the energy management system of the integrated photovoltaic, energy storage and charging device in this invention; Figure 2 This is a detailed flowchart of the energy management system of the integrated photovoltaic, energy storage and charging device in this invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses an energy management system for an integrated photovoltaic, energy storage, and charging device, such as... Figures 1 to 2 As shown, it includes the following steps: The data acquisition module collects multi-source heterogeneous raw data streams, performs preprocessing and state feature extraction, and obtains a standard operating state feature vector. Step 1.1, Distributed sensor network data acquisition; Based on a distributed sensor network deployed across photovoltaic arrays, energy storage systems, charging pile clusters, and grid connection points, and employing Industrial Ethernet and Modbus TCP communication protocols, electrical and environmental parameters at each measurement point are synchronously collected at a sampling period of 200 milliseconds. This yields a multi-source heterogeneous raw data stream, including photovoltaic array output voltage, output current, DC bus voltage, photovoltaic module surface temperature, and ambient irradiance; terminal voltage, charging / discharging current, individual cell temperature, state of charge (SOC), and state of health (SOH) of each battery cluster in the energy storage system; connection status, output voltage, output current, requested charging power, remaining vehicle battery capacity, and estimated charging time of each charging gun in the charging piles; and three-phase voltage, three-phase current, frequency, active power, and reactive power on the grid side. Based on the collected multi-source heterogeneous raw data stream, after preliminary processing through a field edge computing gateway, it is uploaded to the main control unit of the energy management system via a fiber optic communication link, resulting in the multi-source heterogeneous raw data stream.

[0021] Step 1.2, Multi-source data time synchronization processing; Based on the multi-source heterogeneous raw data stream output in step 1.1, a timestamp alignment algorithm is used to synchronize the data from different data sources, mapping all data onto the system's standard time axis to eliminate time deviations caused by acquisition and transmission delays. During the timestamp alignment process, linear interpolation is used to fill in missing data points, and the Laida criterion is used to identify and remove outlier data points exceeding three times the standard deviation, ensuring data continuity and reliability, resulting in a time-aligned multi-source data sequence.

[0022] Step 1.3, data normalization processing; Based on the time-aligned multi-source data sequence output in step 1.2, a sliding window-based data normalization method is used to perform dimensionless transformation on parameters with different physical dimensions. Specifically, for power parameters such as photovoltaic output power, energy storage charging and discharging power, and charging pile load power, their respective rated power is used as the reference value for normalization; for state parameters such as energy storage SOC, voltage, and current, the maximum and minimum values ​​within their normal operating range are used for linear normalization; and for environmental parameters such as temperature and irradiance, the mean and standard deviation of historical statistical data are used for standardization, resulting in a normalized data sequence with values ​​uniformly between 0 and 1.

[0023] Step 1.4, wavelet transform noise reduction processing; Based on the normalized data sequence output in step 1.3, wavelet transform is used to decompose the data into multi-scale components, separating the original signal into high-frequency detail components and low-frequency trend components. Soft thresholding denoising is then used to filter out high-frequency noise while retaining the low-frequency signal reflecting the system's operating trend. Wavelet transform denoising is performed on the normalized data sequence using the Daubechies wavelet as the basis function. The decomposition level is set to three levels, and an adaptive threshold estimation method is used to dynamically determine the denoising threshold based on the signal energy distribution, resulting in the denoised data sequence.

[0024] Step 1.5, accurate estimation of energy storage SOC; Based on the denoised data sequence output in step 1.4, the Extended Kalman Filter (EKF) algorithm is used to accurately estimate the State of Charge (SOC) of the energy storage system. Specifically, SOC estimation is performed using the denoised data sequence by establishing a second-order equivalent circuit model of the energy storage battery. This model includes an open-circuit voltage source, an ohmic internal resistance, and two parallel RC branches, representing the battery's transient and dynamic response characteristics, respectively. Based on this equivalent circuit model, a state-space equation is established, using SOC and the polarization voltages of the two RC branches as state variables, and the terminal voltage as the observation. The EKF algorithm is then used to recursively update the estimated SOC value. This method, by fusing current integral and voltage observation information, can effectively compensate for current sensor drift and initial SOC error, improving the accuracy of SOC estimation and obtaining a corrected energy storage SOC value.

[0025] Step 1.6, Principal component analysis feature extraction; Based on the denoised data sequence output in step 1.4 and the corrected energy storage SOC value output in step 1.5, principal component analysis (PCA) is used to extract the main features of the system's operating state, reducing the data dimensionality. Feature extraction is performed based on the denoised data sequence and the corrected energy storage SOC value. Specifically, key parameters such as photovoltaic output power, energy storage SOC, energy storage charging and discharging power, total charging pile load power, grid interaction power, ambient irradiance, and ambient temperature are combined into a high-dimensional state vector. PCA is used to calculate the covariance matrix of each parameter, and the top principal components with a cumulative contribution rate exceeding 95% are extracted. The high-dimensional state vector is then projected onto a low-dimensional principal component space to obtain a standard operating state feature vector containing key information about system operation, which serves as the input for subsequent prediction and optimization modules.

[0026] Step 1.7, Multi-timescale feature construction; Based on the standard operating state feature vector output in step 1.6, a multi-time-scale feature sequence is constructed using a time-series feature extraction method, resulting in a composite feature representation reflecting the system's short-term fluctuations and long-term trends. Specifically, a sliding time window mechanism is used to extract feature statistics from the most recent 5 minutes, 30 minutes, 2 hours, and 24 hours, including mean, standard deviation, maximum value, minimum value, and rate of change, forming a multi-scale time-series feature set. This multi-scale feature set can simultaneously capture the system's instantaneous state and historical evolution patterns, providing rich input information for the prediction module and yielding a multi-time-scale feature sequence.

[0027] In some embodiments, since the charging load at highway service areas differs between weekdays and holidays, a date-type-based feature enhancement method can be employed. This method adds date-type identifiers to the standard operating state feature vector output in step 1.6, including category labels such as weekday, weekend, and statutory holiday, as well as time location identifiers such as whether it is the day before or the last day of a holiday. The aim is to enhance the prediction module's ability to identify load patterns under different date types and improve prediction accuracy. Specifically, date-type feature enhancement based on the standard operating state feature vector involves establishing a one-hot encoded representation of the date-type labels, converting the date information into a binary vector, and concatenating it with the standard operating state feature vector to form an extended feature vector. By introducing date-type features, the prediction model can automatically learn the load characteristics under different date types, adjust the prediction strategy in advance before holidays, provide more accurate load prediction information to the energy dispatch module, and obtain the extended standard operating state feature vector.

[0028] The prediction module, based on the feature vector of standard operating state, performs intelligent prediction of photovoltaic output and charging load, and obtains intelligent prediction results; Step 2.1, Construction of photovoltaic prediction model through transfer learning; Based on the standard operating state feature vector output in step 1.6, a photovoltaic (PV) power output prediction model is constructed using transfer learning to obtain a multi-timescale PV power generation prediction sequence. Specifically, the PV prediction model is constructed by first collecting historical operating data from multiple geographically similar PV-storage-charging systems with similar installed capacities as the source domain dataset. This source domain dataset contains at least one year of PV power output records and corresponding meteorological data. Based on this source domain dataset, a basic PV power output prediction model is constructed using a Long Short-Term Memory (LSTM) network. The network structure includes an input layer that receives a 128-dimensional standard operating state feature vector; a first-layer LSTM containing 128 neurons that processes the input feature sequence and outputs a hidden state vector; a second-layer LSTM containing 64 neurons that processes the upper-layer output through forgetting, input, and output gates; a third-layer LSTM containing 32 neurons that further extracts temporal features; and a fully connected output layer that maps the final hidden state to a PV power prediction value. Data selectively remembers and forgets data between layers through a gating mechanism. The training process uses mean squared error as the loss function, calculates gradients to update network parameters using the backpropagation algorithm, and optimizes parameters using the Adam optimizer with a learning rate of 0.001. The training cycle is 100 epochs, with performance evaluated on the validation set every 10 epochs and an early stopping mechanism to prevent overfitting. After the basic model training is complete, data collected from the local photovoltaic energy storage and charging system is used as the target domain data. The parameters of the first two layers of the LSTM network are fixed, and only the third LSTM layer and the output layer are fine-tuned. The learning rate is reduced to 0.0001, and the fine-tuning cycle is 20 epochs. Through transfer learning, the newly built system can quickly establish a high-precision prediction model even with insufficient data accumulation, resulting in a photovoltaic power output prediction model.

[0029] Step 2.2, multi-timescale rolling forecast; Based on the photovoltaic power output prediction model output in step 2.1 and the current standard operating state feature vector output in step 1.6, a multi-step rolling prediction method is adopted to generate photovoltaic power prediction values ​​for four time scales: 15 minutes, 1 hour, 4 hours, and 24 hours. Multi-time scale predictions are performed based on the photovoltaic power output prediction model and the current standard operating state feature vector. The 15-minute ultra-short-term prediction mainly relies on the changing trends of current photovoltaic output power and irradiance, using a 5-minute prediction step size and rolling prediction for 3 steps. The 1-hour short-term prediction combines current status and weather forecast information, using a 15-minute prediction step size and rolling prediction for 4 steps. The 4-hour medium-term prediction and the 24-hour long-term prediction mainly rely on meteorological forecast data and historical photovoltaic power output patterns, using a 1-hour prediction step size and rolling prediction for 4 and 24 steps respectively. The prediction results at different time scales provide the energy dispatch module with a full-time-domain decision-making basis from real-time control to day-ahead planning, resulting in multi-time-scale photovoltaic power output prediction curves.

[0030] Step 2.3, Quantile regression uncertainty estimation; Based on the multi-timescale photovoltaic (PV) power output prediction curves output in step 2.2, the quantile regression method is used to estimate the prediction uncertainty, obtaining PV power output prediction intervals at different confidence levels. Specifically, in the output layer of the LSTM prediction model, the quantile loss function is used instead of the traditional mean squared error loss function, training three prediction models with prediction quantiles of 0.1, 0.5, and 0.9, corresponding to the lower prediction limit, median prediction, and upper prediction limit, respectively. Quantile prediction quantifies the prediction uncertainty of PV power output, providing a basis for subsequent risk constraint optimization. In energy dispatch decisions, when a conservative strategy is required, the 0.1 quantile prediction value is used as the available PV power; when an aggressive strategy is required, the 0.9 quantile prediction value is used; and under normal conditions, the median prediction value of the 0.5 quantile is used to obtain the PV power output prediction interval.

[0031] Step 2.4, charging load probability modeling; Based on the historical charging pile load data in the standard operating state feature vector output in step 1.6, a nonparametric kernel density estimation method is used to establish a charging load probability model to obtain the charging load probability density distribution for future periods. Probabilistic modeling is performed based on the historical charging pile load data in the standard operating state feature vector. Specifically, firstly, statistical analysis is conducted on historical charging records to extract the charging pile occupancy rate and average charging power for different time periods and date types. A vehicle arrival model is established based on the Poisson process assumption, with the arrival rate parameter dynamically adjusted according to the time period characteristics of historical statistical data: low arrival rate on weekday mornings and high arrival rate in the afternoons and evenings of holidays. For the charging demand of each arriving vehicle, the Gaussian kernel density estimation method is used to fit the probability distribution of charging power and charging duration based on historical charging record data. The vehicle arrival model and the single-vehicle charging demand model are combined, and load aggregation is performed using Monte Carlo random sampling to generate 1000 sets of future charging load evolution samples. The probability density distribution curve of the charging load and the load intervals at different confidence levels are statistically obtained, thus yielding the charging load probability density distribution.

[0032] Step 2.5, Scene Reduction and Representative Scene Generation; Based on the multi-timescale photovoltaic output prediction curves output in step 2.2 and the charging load probability density distribution output in step 2.4, a scenario reduction technique is used to generate a representative set of operating scenarios, resulting in a typical scenario sequence for optimized scheduling. Scenario reduction is performed based on the multi-timescale photovoltaic output prediction curves and the charging load probability density distribution. Specifically, different prediction quantiles of photovoltaic output are combined with different probability levels of charging load to initially generate a scenario tree containing multiple possible operating scenarios. Since an excessive number of scenarios would drastically increase the computational complexity of the optimization problem, a scenario reduction algorithm based on probability distance is used to retain representative scenarios with higher probabilities and greater differences, compressing the number of scenarios to approximately 10. Each representative scenario corresponds to a photovoltaic output curve, a charging load curve, and an occurrence probability value. These scenarios collectively describe the uncertainty space of the system's future operation, providing input for robust optimized scheduling, resulting in a typical scenario sequence.

[0033] Step 2.6, Net load forecasting and grid interaction demand analysis; Based on the photovoltaic output prediction curve output in step 2.2, the charging load probability distribution output in step 2.4, and the current energy storage SOC state output in step 1.5, the system net load prediction curve is calculated using the energy balance analysis method, resulting in a power prediction sequence that requires purchasing or selling electricity to the grid. Net load prediction is performed based on the photovoltaic output prediction curve, the charging load probability distribution, and the current energy storage SOC state. Specifically, at each future time point, the net load equals the charging load prediction value minus the photovoltaic output prediction value, and then minus the adjustable power of the energy storage. When the net load is positive, it indicates a power deficit in the system, requiring energy storage discharge or purchasing electricity from the grid; when the net load is negative, it indicates a power surplus in the system, allowing for energy storage charging or selling electricity to the grid. Considering the SOC constraints and charging / discharging power limitations of the energy storage, the adjustable power range of the energy storage system in future time periods is calculated, further refining the net load prediction curve to obtain the final grid interaction power demand prediction. This serves as the basis for formulating the day-ahead dispatch plan, resulting in the grid interaction power demand prediction.

[0034] In some embodiments, since the accuracy of weather forecast data decreases with the extension of the forecast period, relying solely on weather forecasts may lead to significant errors in 24-hour long-term photovoltaic (PV) output forecasting. A multi-model ensemble forecasting method can be employed, simultaneously establishing a physical model forecast based on numerical weather prediction and a statistical model forecast based on historical data. The two forecasts are then fused using a Bayesian model averaging method. The aim is to comprehensively utilize the mechanistic nature of the physical model and the data-driven nature of the statistical model to improve the accuracy and robustness of long-term forecasts. Specifically, the physical model calculates power generation using an engineering model of PV cells based on irradiance data from numerical weather prediction and the physical parameters of the PV array. The statistical model establishes a mapping relationship between input features and PV power based on historical PV output data for the same period using a support vector machine regression method. For the forecasts of the two models, their respective prediction variances are estimated based on their historical prediction errors. A weighted average is then performed using the inverse of the variance as the weight, with the model with the smaller prediction error receiving a higher weight. This results in a fused PV output forecast, which exhibits better accuracy and stability than a single model.

[0035] In some embodiments, due to the randomness of charging load being influenced by complex factors such as vehicle travel patterns and driver behavior habits, probability models based solely on historical statistics are insufficient to capture load anomalies caused by special events. A charging load prediction method integrating real-time traffic information can be adopted. This involves calling the real-time traffic data interface of the highway traffic management system to obtain information such as traffic flow, vehicle speed, and congestion index of the upstream road segment of the service area. The aim is to predict the number of vehicles arriving at the service area in the future based on traffic flow trends, thereby improving the foresight of charging load prediction. Specifically, a dynamic correlation model between traffic flow and service area vehicle arrival rate is established. When an increase in traffic flow is detected in the upstream road segment, the vehicle arrival rate parameter for the future period is increased accordingly. When traffic congestion is detected, the impact of extended vehicle travel time on the arrival time distribution is considered, and the time distribution characteristics of the load prediction are dynamically adjusted. By integrating real-time traffic information into the charging load prediction model, the system can identify load growth trends in advance before the arrival of high traffic volumes during holidays, providing the energy scheduling module with more preparation time and ensuring stable charging services during peak load periods.

[0036] This module outputs intelligent prediction results, including photovoltaic power output prediction models, photovoltaic power output prediction curves at multiple time scales, photovoltaic power output prediction intervals, probability density distribution of charging loads, typical scenario sequences, and power demand predictions for grid interaction, providing comprehensive prediction data support for energy optimization and scheduling.

[0037] The optimization scheduling module performs multi-objective energy optimization scheduling based on intelligent prediction results and generates real-time scheduling instructions; Step 3.1, Construction of multi-objective optimization model; Based on the multi-timescale photovoltaic output prediction curves, charging load probability density distribution, and the current energy storage system SOC and SOH parameters output in step 1.5 from the intelligent prediction results output by the prediction module, a multi-objective optimization model is constructed, including economic, environmental, and equipment lifespan objectives. Specifically, the economic objective function is to minimize the system's total daily operating cost, including the cost of purchasing electricity from the grid, revenue from selling electricity to the grid, and energy storage loss costs. The electricity purchase cost equals the sum of the products of the purchased power and the corresponding time-of-use electricity price for each time period; the electricity sales revenue equals the sum of the products of the sold power and the grid-connected electricity price for each time period; and the energy storage loss cost is related to the energy storage's charging and discharging capacity and its current health status. The environmental objective function is to maximize the photovoltaic absorption rate, i.e., to maximize the ratio of the actual daily photovoltaic power utilization to the total predicted photovoltaic power generation. The objective function for equipment lifespan is to minimize the equivalent cycle number of the energy storage system. The charge-discharge cycle depth of the energy storage SOC is statistically determined by the rainflow counting method, and the equivalent full cycle number is calculated based on the energy storage lifespan decay model as a quantitative indicator of lifespan loss, thus obtaining a comprehensive optimization framework.

[0038] Step 3.2, Construction of the adaptive weighted comprehensive objective function; Based on the comprehensive optimization framework output in step 3.1, a normalization method is used to transform the three objective functions to a unified dimension, and adaptive weight coefficients are introduced to construct a weighted sum form of the comprehensive objective function, resulting in a single-objective optimization problem. Adaptive weight settings are implemented based on the comprehensive optimization framework. Specifically, the baseline weights for the three objectives are set according to the operator's preferences: the economic objective baseline weight is set to 0.5, the environmental objective baseline weight is set to 0.3, and the equipment lifespan objective baseline weight is set to 0.2. In actual operation, the weight coefficients are dynamically adjusted based on the current energy storage health status and electricity price level. When the energy storage health status is good, the weight of the equipment lifespan objective is appropriately reduced, allowing the energy storage to undertake more regulation tasks to improve economic efficiency; when the energy storage health status declines, the weight of the equipment lifespan objective is increased to reduce the frequency of energy storage use and delay degradation. When the peak-valley difference in electricity prices is large, the weight of the economic objective is increased to fully utilize peak-valley arbitrage opportunities; when the electricity price is stable, the weight of the environmental objective is increased to prioritize the consumption of photovoltaic power generation, resulting in a single-objective optimization problem.

[0039] Step 3.3: Establish the set of constraints; Based on the single-objective optimization problem output in step 3.2, a set of constraints for day-ahead energy dispatch is established, resulting in a complete day-ahead optimal dispatch mathematical model. The constraints, established based on the single-objective optimization problem, include: power balance constraints, meaning that at any given time, the sum of photovoltaic power generation, energy storage discharge power, and grid-purchased power equals the sum of charging load power, energy storage charging power, and grid-sold power; energy storage SOC constraints, limiting SOC operation to between 20% and 90% to prevent overcharging and over-discharging damage to the batteries; energy storage charging and discharging power constraints, ensuring that charging and discharging power does not exceed the rated power of the energy storage converter; grid interaction power constraints, ensuring that both grid-purchased and grid-sold power do not exceed the grid connection capacity of 400 kW; charging pile output power constraints, ensuring that the output power of each charging pile does not exceed its rated power of 120 kW; and mutual exclusion constraints for energy storage charging and discharging states, preventing simultaneous charging and discharging of energy storage. By introducing 0-1 integer variables to represent the charging and discharging states, a complete day-ahead optimal dispatch mathematical model is obtained.

[0040] Step 3.4, Solve the mixed-integer linear programming problem; Based on the complete day-ahead optimal scheduling mathematical model output in step 3.3, a mixed-integer linear programming algorithm is used to solve the problem, obtaining the optimal scheduling plan for energy storage charging and discharging power, grid interaction power, and photovoltaic utilization power for each time period in the next 24 hours. Specifically, the commercial optimization solver CPLEX or the open-source solver GLPK are used to solve the model, with a solution time limit of 10 minutes and a relative error tolerance of 1%. During the solution process, nonlinear constraints are first linearized, and the mutual exclusion constraint for energy storage charging and discharging is converted into a linear inequality constraint using the Big M method. The obtained optimal scheduling plan, with a time resolution of 1 hour, provides the power allocation scheme for each time period. This plan serves as a benchmark reference for system operation, guiding the execution of real-time scheduling and yielding the optimal scheduling plan.

[0041] Step 3.5, Real-time rolling optimization scheduling; Based on the optimal scheduling plan output in step 3.4 and the real-time updated standard operating state feature vector output in step 1.6, a rolling optimization strategy is used to establish a real-time energy scheduling model, obtaining refined power allocation instructions for the next hour. Real-time rolling optimization is performed based on the optimal scheduling plan and the real-time updated standard operating state feature vector. Specifically, real-time scheduling is performed in 15-minute time intervals, with the optimization window covering four time intervals totaling one hour, and optimization is performed every 15 minutes. The objective function and constraints of the real-time scheduling model are basically the same as those of the day-ahead scheduling model, but a deviation penalty term for tracking the day-ahead baseline plan is added to prevent excessive deviation between real-time scheduling and the day-ahead plan. A simplified linear programming model is used for real-time scheduling, removing some non-critical constraints and controlling the solution time to within 10 seconds, meeting the response requirements of real-time control. The real-time scheduling model can correct day-ahead prediction deviations based on the actual values ​​of photovoltaic output and charging load, dynamically adjusting the power allocation scheme to ensure the system maintains its optimal state during actual operation, obtaining refined power allocation instructions.

[0042] Step 3.6, Power allocation command generation; Based on the refined power allocation instructions output in step 3.5, the power allocation instructions for the current time period are extracted, including the set value of energy storage charging or discharging power, the set value of grid power purchase or sales power, and the power allocation coefficient of each charging pile. Real-time scheduling instructions are then sent to the underlying controller, which is the hardware device that actually executes the power control instructions, including photovoltaic inverters, energy storage converters, and charging pile controllers.

[0043] Command generation is based on refined power allocation instructions. Specifically, the energy storage power instruction includes a target power value and a power ramp-up rate limit. The ramp-up rate is set according to the dynamic response capability of the energy storage converter, with the charging ramp-up rate limited to 20% of the rated power per second and the discharging ramp-up rate limited to 30% of the rated power per second. Grid interaction power instructions must meet the power change rate requirements of the grid connection point, with power changes not exceeding 10% of the connected capacity per unit time to avoid impacting the grid. The charging pile power allocation coefficient is dynamically allocated based on the current charging demand of each charging pile and the vehicle battery status, prioritizing vehicles with low remaining battery power and early expected departure times. When total power is limited, power is reduced according to priority, resulting in real-time dispatch instructions.

[0044] In some embodiments, since the cycle life of energy storage batteries is closely related to their charge / discharge depth and charge / discharge rate, frequent deep charge / discharge and high-rate operation will accelerate battery degradation. A scheduling strategy based on energy storage health status hierarchical management can be adopted, which divides the energy storage into four health levels: excellent, good, average, and declining, according to the current SOH level. The purpose is to adopt differentiated scheduling strategies for energy storage units with different health levels, with units in good health undertaking more regulation tasks and units in poor health having reduced usage intensity, thereby achieving full life cycle optimization management of energy storage assets. Specifically, in the optimized scheduling model, different constraints are set for energy storage units with different health levels: For units with an excellent health level, the SOC operating range is 20% to 90%, and the charging / discharging power can reach 100% of the rated power; for units with a good health level, the SOC operating range is narrowed to 30% to 80%, and the charging / discharging power is limited to 80% of the rated power; for units with a moderate health level, the SOC operating range is further narrowed to 40% to 70%, and the charging / discharging power is limited to 60% of the rated power; for units with a declining health level, they only participate in scheduling when the system experiences a power deficit and other energy storage units are insufficient, with an SOC operating range of 45% to 65% and a charging / discharging power limited to 40% of the rated power. Through graded management of health status, while ensuring the system's regulation capability, the overall degradation rate of the energy storage battery is effectively slowed down, the service life of the energy storage system is extended, and the total life cycle cost is reduced.

[0045] In some embodiments, due to the significant differences in grid time-of-use prices during peak and off-peak periods, rationally utilizing these prices for energy storage arbitrage can reduce system operating costs. An optimization method for energy storage charging and discharging strategies considering the characteristics of different electricity price periods can be adopted. During off-peak periods, priority is given to charging energy storage with low-priced grid electricity, while during peak periods, priority is given to releasing energy storage power to avoid purchasing electricity at high prices. The aim is to reduce the system's electricity purchase costs from the grid and improve economic efficiency through the peak-shaving and valley-filling effects of energy storage. Specifically, in the day-ahead dispatch model, soft constraints based on electricity price periods are introduced to encourage energy storage to charge during the low-price period from 0:00 to 6:00 AM and discharge during the high-price periods from 10:00 to 12:00 PM and from 5:00 to 9:00 PM. To avoid sacrificing photovoltaic (PV) absorption and equipment lifespan in pursuit of excessive economic goals, hard constraints are set, including a PV absorption rate of no less than 85% and a daily equivalent cycle count of no more than 0.5 times. Through electricity price period optimization, the system can fully utilize price differences to obtain arbitrage profits while meeting charging service needs and PV absorption targets, thereby reducing daily operating costs to a more optimal level and improving economic efficiency.

[0046] The real-time control module performs equipment power control and coordination operations based on real-time scheduling instructions, and obtains equipment power control results. Step 4.1: Issuance of standardized communication protocol instructions; Based on the real-time scheduling instructions output in step 3.6, power commands are sent to the photovoltaic inverters, energy storage converters, and charging pile controllers using a standardized industrial communication protocol, obtaining feedback signals confirming receipt from each device. Specifically, the energy management system's main control unit establishes communication connections with each underlying controller via an industrial Ethernet switch, using the Modbus TCP protocol for data exchange. Instructions sent to the photovoltaic inverters include active power setpoints and power factor setpoints. The active power setpoint indicates the current power level that the photovoltaic system should output. When it is necessary to limit photovoltaic output, power limiting is achieved by adjusting the operating point of the maximum power point tracking algorithm. Instructions sent to the energy storage converters include operating mode selection, active power setpoints, and reactive power setpoints. Operating modes include constant power charging, constant power discharging, and standby mode. Instructions sent to each charging pile controller include a charging enable signal and a power allocation coefficient. The power allocation coefficient multiplied by the rated power of the charging pile yields the currently allowed maximum output power. All instructions are transmitted using the highest real-time priority communication task, with end-to-end communication latency controlled within 50 milliseconds, and feedback signals confirming receipt are obtained from each device.

[0047] Step 4.2, droop control of energy storage converter; Based on the power setpoint received by the energy storage converter in the feedback signals confirmed by each device in step 4.1, a power regulation algorithm based on droop control is used to control the charging and discharging behavior of the energy storage system, obtaining the DC bus voltage and the actual charging and discharging power that accurately tracks the setpoint. Droop control is performed based on the power setpoint received by the energy storage converter. Specifically, the energy storage converter adopts a dual closed-loop control structure. The outer loop is the power control loop, which calculates the current reference value through a proportional-integral controller (PI) based on the deviation between the power setpoint and the actual power. The inner loop is the current control loop, which calculates the modulation signal through a PI controller based on the deviation between the current reference value and the actual current, driving the switching of power devices. The DC bus voltage is stabilized through the voltage outer loop. When multiple energy storage converters are operating in parallel, a droop control strategy is used to achieve automatic power distribution. Each converter is set with a different droop coefficient according to its capacity and health status; converters with larger capacity and better health status receive more power. The control cycle of the energy storage converter is set to 100 microseconds, which enables rapid dynamic response. It can adjust the output power in a timely manner when the load changes abruptly, maintain the power balance of the system, and obtain the DC bus voltage and the actual charging and discharging power.

[0048] Step 4.3, adaptive MPPT control of photovoltaic inverter; Based on the power setpoint received by the photovoltaic inverter from the feedback signals confirmed by each device in step 4.1, an adaptive maximum power point tracking (MPPT) algorithm is used to adjust the operating point of the photovoltaic system to obtain the actual photovoltaic output power that matches the system requirements. Adaptive MPPT control is performed based on the power setpoint received by the photovoltaic inverter. Specifically, the photovoltaic inverter has a built-in MPPT algorithm. Under normal circumstances, the MPPT algorithm drives the photovoltaic array to operate at its maximum power point, maximizing the utilization of photovoltaic energy. When the power setpoint issued by the energy management system is lower than the current maximum power of the photovoltaic array, the inverter switches to constant power control mode, offsetting the operating point to make the photovoltaic output power equal to the setpoint. At this time, some photovoltaic energy is not utilized. To reduce photovoltaic curtailment, photovoltaic output is only limited when the energy storage is fully charged and the charging load is insufficient; otherwise, photovoltaic power generation is prioritized. The photovoltaic inverter also has low voltage ride-through and frequency regulation functions. When the grid voltage drops or the frequency is abnormal, the inverter does not disconnect from the grid but provides reactive power support or active power regulation, enhancing the grid friendliness of the system and obtaining the actual photovoltaic output power.

[0049] Step 4.4, constant current and constant voltage control of the charging pile; Based on the power allocation coefficients received by each charging pile controller in the feedback signals confirmed by each device in step 4.1, a constant current and constant voltage two-stage charging control strategy is adopted to adjust the output power of the charging piles, thereby obtaining a stable charging output that meets the charging needs of electric vehicles. Constant current and constant voltage control is performed based on the power allocation coefficients received by each charging pile controller. Specifically, the charging pile controller determines the current charging stage based on the battery status information fed back by the vehicle's Battery Management System (BMS). In the constant current charging stage, the charging pile outputs a constant current, the value of which is equal to the maximum power corresponding to the power allocation coefficient divided by the current battery voltage, achieving maximum power charging. In the constant voltage charging stage, the charging pile outputs a constant voltage, and the charging current gradually decreases as the battery's state of charge increases, naturally reducing the charging power. The charging pile controller monitors the output voltage, current, and power in real time to ensure that they do not exceed the limits set by the power allocation coefficient. When the energy management system issues a power reduction command, the charging pile reduces its output power according to priority, prioritizing vehicles that are about to be fully charged or are expected to leave early, while appropriately extending the charging time for other vehicles to achieve a stable charging output.

[0050] Step 4.5, Power Deviation Detection and Dynamic Compensation; Based on the actual operating status feedback data of each execution unit, including the DC bus voltage and actual charging / discharging power output in step 4.2, the actual photovoltaic output power output in step 4.3, and the stable charging output output in step 4.4, a power deviation detection and dynamic compensation algorithm is used to obtain the corrected power control command. Power deviation detection is performed based on the actual operating status feedback data of each execution unit. Specifically, the energy management system collects the actual output power of the photovoltaic inverter, energy storage converter, and charging pile in real time and compares it with the issued power setpoint to calculate the power tracking deviation. When the deviation exceeds the set threshold, the cause of the deviation is analyzed, which may be due to equipment response delay, line loss, or equipment failure. For response delay, the response speed is accelerated by adding a feedforward control component; for line loss, the power setpoint is corrected according to the loss model; for equipment failure, a fault isolation procedure is triggered and the power command is redistributed. A composite control structure of feedforward and feedback is adopted. Feedforward control adjusts the equipment status in advance based on predicted information, while feedback control performs precise correction based on real-time deviation. The combination of the two achieves high-precision power tracking control, resulting in the corrected power control command.

[0051] Step 4.6, Power ramp rate limiting and grid-friendly control; Based on the stable operating states of each execution unit, including the DC bus voltage and actual charging / discharging power output in step 4.2, the actual photovoltaic output power output in step 4.3, and the stable charging output output in step 4.4, a power ramp-up rate limiting algorithm is used to control the rate of change of grid interaction power, resulting in a smooth power curve that meets grid connection requirements. Power ramp-up rate limiting is implemented based on the stable operating states of each execution unit. Specifically, the target change in grid interaction power for each control cycle is calculated and compared with the maximum allowable ramp-up rate. When the target change exceeds the ramp-up limit, the actual power change is limited to the allowable range, and any unfinished power adjustments are postponed to the next cycle. This power ramp-up limiting prevents sudden changes in photovoltaic output and load from being directly transmitted to the grid, effectively reducing the impact on the distribution network and improving the grid-friendliness of the system. When grid frequency or voltage anomalies are detected, the energy management system automatically triggers an emergency response mode, adjusting the control strategies of the photovoltaic inverter and energy storage converter to provide frequency or voltage support, participating in grid stability control, and obtaining a smooth power curve.

[0052] In some embodiments, circulating currents may occur when multiple energy storage converters operate in parallel, increasing system losses and affecting energy storage lifespan. An improved virtual synchronous machine control strategy can be employed, assigning virtual inertia and damping characteristics to each energy storage converter to simulate the external characteristics of a synchronous generator. The aim is to achieve self-synchronization between energy storage converters through virtual synchronous machine control, avoiding circulating currents and improving the system's inertial support capability. Specifically, in the control algorithm of the energy storage converter, a virtual rotor motion equation is introduced. A virtual angular frequency is calculated based on the active power deviation, and a virtual voltage amplitude is calculated based on the reactive power deviation. Power decoupling is achieved through virtual impedance. The virtual inertia parameters of each converter are configured according to its capacity; larger-capacity converters are assigned larger virtual inertia to provide more inertial support during power disturbances. Virtual damping parameters are used to suppress power oscillations; by adjusting the damping coefficient, a balance can be achieved between dynamic response speed and stability margin. By adopting virtual synchronous machine control, the energy storage converter can not only accurately track power commands, but also has frequency and voltage regulation capabilities similar to synchronous generators, thereby improving the grid support capability and disturbance resistance capability of the photovoltaic-storage-charging system.

[0053] In some embodiments, since charging load may exceed the system's power supply capacity during peak holiday periods, directly rejecting some charging requests would negatively impact user experience. A dynamic power allocation strategy based on flexible charging demand management can be adopted. This strategy obtains information on charging time flexibility through user interaction and offers electricity price discounts to users who are not in a hurry to charge, guiding them to postpone their charging. The aim is to smooth out load peaks and improve the system's load acceptance capacity while ensuring necessary charging services. Specifically, charging piles are equipped with a human-machine interface, allowing users to choose between immediate charging mode or economy charging mode before starting charging. Users choosing immediate charging mode are charged at the normal electricity price, and the system prioritizes their charging needs. Users choosing economy charging mode enjoy electricity price discounts, but their charging time may be delayed; the system schedules their charging during off-peak hours. The energy management system dynamically adjusts the price difference between the two modes based on the current load level and energy storage status. When load pressure is high, the electricity price discount is increased to attract users to choose economy mode. By guiding user behavior through price signals, flexible adjustment of charging load is achieved. During peak holiday periods, the proportion of users choosing economy charging mode can reach over 30%, effectively alleviating the system's power supply pressure and improving overall service quality.

[0054] This module outputs the power control results of the devices, including the feedback signals received and confirmed by each device, the DC bus voltage and the actual charging and discharging power, the actual output power of the photovoltaic system, the stable charging output, the corrected power control command and the smoothed power curve, thus realizing the coordinated control and precise power tracking of photovoltaic, energy storage and charging pile devices.

[0055] The health management module performs energy storage health status assessment and system performance analysis based on equipment power control results and standard operating status feature vectors, and obtains energy storage health status assessment and health management results. Step 5.1, Extraction of features from energy storage operation data; Based on the energy storage system operation data continuously output by the data acquisition module from the standard operating state feature vector, including parameters such as terminal voltage, charging and discharging current, individual cell temperature, and SOC of each battery cluster, data preprocessing and feature extraction methods are used to obtain a set of key feature parameters reflecting the energy storage operation status. Feature extraction is performed based on the energy storage system operation data. Specifically, for the voltage data of each battery cluster, its average value, standard deviation, maximum value, and minimum value are calculated. The voltage standard deviation reflects the consistency of individual battery cells; an increase in the standard deviation indicates an imbalance within the battery pack. For the charging and discharging current data, the distribution of charging and discharging rates is statistically analyzed; excessively high rates accelerate battery aging. For the temperature data, abnormal temperature rise events are identified; excessively high temperatures are an important sign of potential battery safety hazards. These feature parameters are organized into a time series to form an energy storage operation status feature database, resulting in the set of key feature parameters.

[0056] Step 5.2, Rainflow counting cycle statistics; Based on the cumulative charge-discharge records of the energy storage system from the DC bus voltage and actual charge-discharge power output in step 4.2, the rainflow counting method is used to statistically analyze the charge-discharge cycles of the energy storage SOC, obtaining the cycle number distribution at different depths. Rainflow counting is performed based on the cumulative charge-discharge records of the energy storage system. Specifically, the rainflow counting method is a cycle counting method widely used in fatigue life analysis, capable of extracting complete charge-discharge cycles from complex SOC fluctuation curves. The algorithm first identifies the peak and valley values ​​in the SOC sequence, then matches the peak-valley pairs according to the rainflow rule to form cycles. Each cycle is characterized by its cycle depth and average SOC. The cycle depth equals the peak SOC minus the valley SOC, and the average SOC equals the peak SOC plus the valley SOC divided by 2. The cumulative distribution of the cycle number of the energy storage system at different SOC intervals and different cycle depths is statistically obtained, serving as the basic data for health status assessment, thus yielding the cycle number distribution.

[0057] Step 5.3, Calculate the equivalent total number of iterations; Based on the cycle number distribution and battery cycle life curve output in step 5.2, the equivalent total number of cycles of the energy storage system is calculated using the linear cumulative damage theory to obtain an estimate of the energy storage capacity decay. The equivalent total number of cycles is calculated based on the cycle number distribution. Specifically, according to the cycle life data provided by the battery manufacturer, different cycle depths correspond to different cycle lives; for example, 100% depth has a cycle life of 3000 cycles, 50% depth has a cycle life of 6000 cycles, and 25% depth has a cycle life of 12000 cycles. Using the Miner linear cumulative damage criterion, the lifespan damage caused by each actual cycle is equivalent to the total number of cycles. The equivalent total number of cycles caused by a cycle at a certain depth is equal to that cycle number divided by the cycle life corresponding to that depth, multiplied by 3000. The equivalent total number of cycles for all cycles is summed to obtain the current total equivalent total number of cycles for the energy storage system. Based on the relationship curve between the equivalent number of cycles and capacity decay, the current remaining capacity ratio of the energy storage is estimated to obtain an estimate of the energy storage capacity decay.

[0058] Step 5.4, Calculate the overall health score; Based on the set of key characteristic parameters output in step 5.1 and the estimated energy storage capacity decay output in step 5.3, a comprehensive scoring method is used to calculate the energy storage health score, resulting in an energy storage health status index. The health score is calculated based on four dimensions: capacity health, internal resistance health, temperature health, and consistency health. Capacity health is determined by the estimated energy storage capacity decay output in step 5.3; 100% remaining capacity corresponds to a full score of 100 points, 80% remaining capacity corresponds to 80 points, and so on. Internal resistance health is determined by the growth rate of the battery's internal resistance, obtained through periodic static capacity testing; no increase in internal resistance corresponds to a full score, and a 50% increase corresponds to 60 points. Temperature health is determined by the statistical distribution of operating temperature; consistently maintaining the optimal temperature range corresponds to a full score, while frequent high temperatures correspond to a low score. Consistency health is determined by the standard deviation of the battery cell voltage; a smaller standard deviation indicates better consistency. The four dimensions of scores are weighted and averaged with weights of 0.4, 0.3, 0.2, and 0.1 to obtain a comprehensive health score. Based on the score, the energy storage health status is divided into four levels: excellent (above 90 points), good (75 to 90 points), average (60 to 75 points), and declining (below 60 points), thus obtaining a quantitative energy storage health status index.

[0059] Step 5.5, Adjust the adaptive operation strategy; Based on the quantified energy storage health status indicators output in step 5.4, an adaptive strategy adjustment method is used to dynamically correct the operating constraint parameters of the energy storage system, resulting in an energy storage operation strategy. The operation strategy is adjusted based on the quantified energy storage health status indicators. Specifically, for energy storage units with an excellent health level, the original operating constraints are maintained, with a State of Charge (SOC) operating range of 20% to 90%, and charging / discharging power reaching rated power, allowing participation in all scheduling tasks. For energy storage units with a good health level, operating constraints are appropriately tightened, with the SOC operating range adjusted to 30% to 80%, and charging / discharging power limited to 85% of rated power, prioritizing participation in scheduling tasks with shallower charging / discharging depths. For energy storage units with a moderate health level, operating constraints are tightened, with the SOC operating range adjusted to 40% to 70%, and charging / discharging power limited to 70% of rated power, participating in scheduling only when necessary. For energy storage units with a declining health level, usage is strictly limited, with the SOC operating range adjusted to 45% to 65%, and charging / discharging power limited to 50% of rated power, only used as a backup resource when the system experiences a power deficit. The adjusted running constraint parameters are fed back to the optimization scheduling module in step 3.3 to update the constraints of the optimization model and obtain the adaptive running strategy adjustment instruction.

[0060] Step 5.6, Isolated Forest Anomaly Detection; Based on the time-series data of the key feature parameter set output in step 5.1, the isolated forest anomaly detection algorithm is used to identify abnormal operating modes of the energy storage system and obtain early warning signals for potential faults. Anomaly detection is performed based on the time-series data of the key feature parameter set. Specifically, isolated forest is an unsupervised anomaly detection method based on ensemble learning, whose core idea is that abnormal data points are more easily isolated. The algorithm constructs multiple isolation trees by randomly selecting features and segmentation thresholds. For each data point, its average path length in all isolation trees is calculated; data points with shorter path lengths are more likely to be anomalies. The isolated forest model is trained using the energy storage operating status feature parameters as input, and an anomaly score threshold of 0.7 is set. When the anomaly score of the real-time monitoring data exceeds the threshold, an anomaly warning is triggered. The system automatically records the time of the anomaly and the feature parameter value, and sends the warning information to maintenance personnel through the communication interface. Common anomaly modes include: a sudden increase in the voltage difference between individual battery cells, indicating a possible individual cell fault; an abnormal increase in temperature during charging and discharging, indicating a possible internal short circuit; and an increase in internal resistance under the same operating conditions, indicating accelerated battery aging, thus obtaining potential fault warning signals.

[0061] Step 5.7, Expert System Fault Diagnosis; Based on the potential fault warning signals and fault diagnosis rule base output in step 5.6, an expert system method is used to determine the severity and type of the anomaly, and to obtain corresponding protection measures and handling suggestions. Fault diagnosis is performed based on the potential fault warning signals. Specifically, a fault diagnosis rule base is established, containing characteristic descriptions and handling schemes for common fault modes. When an anomaly is detected, the anomaly characteristics are matched with the fault modes in the rule base to determine the fault type and severity. For minor anomalies, such as single-cell voltage imbalance not exceeding safety limits, the system automatically initiates equalization control, adjusting the voltage of each single cell through an active equalization circuit. For moderate anomalies, such as temperature exceeding a safety threshold but not reaching a dangerous level, the system reduces the energy storage charging and discharging power, increases heat dissipation time, and sends a maintenance reminder. For severe anomalies, such as a detected drop in battery cell voltage or a sharp rise in temperature, the system immediately disconnects the charging and discharging circuit of the battery cluster, isolates it from the system, and activates the emergency response plan to ensure system safety. All abnormal events and handling measures are recorded in the system log, providing data support for subsequent fault analysis and system optimization, and obtaining protection measures and handling suggestions.

[0062] Step 5.8, Energy flow tracking and statistics; Based on the energy data records of the system's daily operation, including the DC bus voltage and actual charging / discharging power output in step 4.2, the actual photovoltaic output power output in step 4.3, and the stable charging output output in step 4.4, such as photovoltaic power generation, energy storage charging / discharging, charging pile power supply, and grid interaction power, an energy flow tracking method is used to obtain the power statistics for each energy path in the system. Energy flow tracking is performed based on the energy data records of the system's daily operation. Specifically, an energy flow statistical ledger is established to record the energy source and destination for each time period. Photovoltaic power generation is divided into three destinations: direct photovoltaic power supply to charging piles, photovoltaic charging power stored in energy storage, and photovoltaic power fed into the grid. Energy storage charging sources are divided into photovoltaic charging and grid charging, and energy storage discharging destinations are divided into supplying charging piles and selling electricity to the grid. Charging pile power sources are divided into three parts: direct photovoltaic supply, energy storage power supply, and grid power supply. By statistically analyzing and accumulating data over time periods, the total electrical energy of each energy path throughout the day is obtained. A Sankey diagram of the system's energy flow is then drawn to visually demonstrate the flow and conversion relationships of energy between different units, and the statistical results of the electrical energy of each energy path are obtained.

[0063] Step 5.9, Calculation of key performance indicators; Based on the power statistics of each energy path output in step 5.8, key performance indicator (KPI) calculation methods are used to obtain operational performance evaluation indicators. Specifically, the following core indicators are calculated based on the power statistics of each energy path: Photovoltaic self-consumption rate (equal to the sum of direct photovoltaic power supply and photovoltaic charging power divided by the total photovoltaic power generation, reflecting the local consumption level of photovoltaic energy); energy storage cycle efficiency (equal to the energy storage discharge divided by the energy storage charging power, reflecting the energy conversion efficiency of the energy storage system); photovoltaic power supply ratio of charging piles (equal to the sum of direct photovoltaic power supply and the portion of energy storage discharge from photovoltaic power divided by the total power consumption of charging piles, reflecting the proportion of clean energy used in charging services); peak grid interaction power (reflecting the impact of the system on the grid); and grid interaction power volatility (reflecting power stability by calculating the standard deviation of grid power). Each indicator is compared with the design target value to evaluate whether the system has achieved the expected performance, thus obtaining the operational performance evaluation indicators.

[0064] Step 5.10, Performance Trend Analysis; Based on the long-term statistics of the operational performance evaluation indicators output in step 5.9, trend analysis is used to identify the changing trends of system performance, thereby obtaining performance evolution patterns and potential problem indicators. Specifically, trend analysis is performed based on the long-term statistics of the operational performance evaluation indicators, plotting trend curves of each key indicator over time and analyzing whether the indicators show an upward, downward, or fluctuating trend. If the photovoltaic self-consumption rate shows a downward trend, it may be due to insufficient regulation capacity caused by energy storage capacity decay, or a change in charging load patterns. If the energy storage cycle efficiency continues to decline, it indicates that the energy storage system is aging and needs to be replaced or maintained. If the grid interaction power fluctuation rate increases, it indicates that the system's power smoothing capability is weakening and the control strategy needs to be optimized. Through trend analysis, early detection of signs of system performance degradation provides a basis for decision-making for preventative maintenance and system upgrades, yielding performance evolution patterns and potential problem indicators.

[0065] In some embodiments, since the charging performance of energy storage batteries deteriorates at low temperatures, forced charging may lead to lithium deposition, affecting battery life. A temperature-adaptive charging power control strategy can be employed, which monitors battery temperature in real time. When the temperature falls below a first temperature threshold, the charging power is automatically reduced; when the temperature falls below a second temperature threshold, charging is paused and a heating device is activated. The aim is to protect the battery in low-temperature environments and avoid irreversible damage caused by low-temperature charging. Specifically, the first temperature threshold is set to 5 degrees Celsius, and the second temperature threshold is set to 0 degrees Celsius. When the battery temperature is detected to be below 5 degrees Celsius, the charging power is limited to 50% of the rated power, reducing the charging current to slow the temperature drop. When the temperature falls below 0 degrees Celsius, charging is completely stopped, and the battery thermal management system is activated to preheat the battery, using an electric heating film or heat pump device to raise the battery temperature. When the battery temperature rises back above 5 degrees Celsius, the charging power is gradually restored. For every 1 degree Celsius increase in temperature, the charging power limit is increased by 10%, until the temperature reaches 15 degrees Celsius, at which point the power limit is lifted. Through temperature adaptive control, the energy storage battery can operate safely and reliably under various ambient temperatures, avoiding damage to battery life caused by low-temperature charging.

[0066] This module outputs energy storage health status assessment results and health management results. The energy storage health status assessment results include a set of key characteristic parameters, cycle number distribution, estimated energy storage capacity decay, quantified energy storage health status indicators, and potential fault warning signals. The health management results include energy storage operation strategies, protection measures, power statistics for each energy path, operation performance evaluation indicators, and performance evolution patterns, realizing full life cycle health management of the energy storage system and overall system performance monitoring.

[0067] The strategy learning module builds a case library based on health management results, acquires the characteristics of the operating scenarios, learns and optimizes the control strategies and verifies the operating effects, and obtains intelligent control strategies that evolve autonomously through learning. Step 6.1, run the case library build; Based on the health management results output by the health management module, a case-based structured storage method is used to obtain an operational case library. The case library is constructed based on historical data accumulated over long-term system operation. Specifically, each system operation is recorded as a case, and each case includes three parts: input, process, and output. The input part records the characteristics of the operational scenario, including date type, weather conditions, predicted photovoltaic output, predicted charging load, and initial SOC of energy storage. The process part records control strategy parameters, including optimization target weight coefficients, energy storage charging and discharging strategies, and power allocation schemes. The output part records operational performance indicators, including actual photovoltaic absorption rate, operating costs, energy storage cycle count, and charging service quality. Cases are stored in the case library in chronological order, and a multi-dimensional index is established to support rapid retrieval of similar cases by scenario characteristics, resulting in the operational case library.

[0068] Step 6.2, Similar Case Search; Based on the operational case library output in step 6.1 and the characteristics of the current operational scenario, a case similarity calculation method is used to retrieve historical similar cases from the case library, obtaining a set of reference cases that best match the current scenario. The similarity retrieval is performed based on the operational case library and the characteristics of the current operational scenario. Specifically, a similarity measurement function for scenario features is defined, comprehensively considering multiple dimensions such as date type similarity, meteorological condition similarity, photovoltaic output prediction similarity, and charging load prediction similarity. For categorical features such as date type, similarity is calculated using matching degree, with a perfect match resulting in a similarity of 1 and a non-match resulting in a similarity of 0. For numerical features such as photovoltaic output and charging load, similarity is calculated using Euclidean distance or cosine similarity, with closer numerical values ​​resulting in higher similarity. The comprehensive similarity between the current scenario and each historical case in the case library is calculated; the similarity is a weighted average of the similarities across all dimensions, with weights set according to the importance of the features. The top 10 cases with the highest comprehensive similarity are selected as the reference case set. The control strategies and operational effects of these cases are extracted to provide a reference for the control decisions of the current scenario, resulting in the reference case set.

[0069] Step 6.3, case adaptive strategy generation; Based on the control strategies and operational effects of the reference case set output in step 6.2, an adaptive case-based method is used to generate a control strategy suitable for the current scenario, resulting in a preliminary control parameter scheme. Adaptive strategy generation is performed based on the control strategies and operational effects of the reference case set. Specifically, the best-performing case among the reference cases is analyzed, and its control strategy parameters are extracted as a benchmark scheme. The control parameters are then adjusted according to the differences between the current scenario and the optimal case. For example, if the predicted photovoltaic output of the current scenario is higher than that of the optimal case, the weight of the economic target is appropriately reduced, while the weight of the environmental target is increased to enhance photovoltaic absorption. If the predicted charging load of the current scenario is higher than that of the optimal case, the reserved energy storage capacity is appropriately increased to ensure sufficient regulation capacity during peak load periods. For cases where multiple reference cases have comparable effects, a weighted average method is used to fuse the control strategies of multiple cases, with the weights proportional to case similarity and operational effects. Through case reasoning and adaptive adjustment, a control strategy adapted to the current scenario is generated, avoiding blind exploration from scratch, improving the scientific nature of control decisions, and obtaining a preliminary control parameter scheme.

[0070] Step 6.4, Evaluation of Operational Results and Case Quality Labeling; Based on the operational performance evaluation indicators in the health management results output by the health management module, and awaiting the actual operational results after the current operational cycle ends, the case is scored using an effectiveness evaluation method to obtain a case quality label. The evaluation is based on the actual operational results after the current operational cycle ends. Specifically, the achieved photovoltaic absorption rate, operating costs, energy storage losses, and charging service quality are compared with the expected targets to calculate the achievement degree of each indicator. The photovoltaic absorption rate achievement degree equals the actual absorption rate divided by the target absorption rate; the achievement degrees of other indicators are calculated similarly. A weighted average of the achievement degrees of each indicator is used to obtain the comprehensive score for this operation, ranging from 0 to 100 points. Scores above 90 points are marked as excellent cases, 75 to 90 points as good cases, 60 to 75 points as average cases, and below 60 points as poor cases. The case quality labels are stored in the case library as an important reference for subsequent case retrieval and strategy selection. Excellent cases are prioritized when similarity is the same, while strategies for poor cases are avoided, resulting in the case quality label.

[0071] Step 6.5, Control Strategy Pattern Extraction; Based on the excellent and poor cases in the operational case library output in step 6.1, a comparative analysis method is used to extract the effective and failure modes of the control strategy, thus gaining a regular understanding of strategy optimization. Specifically, excellent and poor cases are grouped according to scenario type, and the differences in control strategies between the two types of cases are compared within the same scenario type. Statistical analysis reveals that excellent cases generally share the following characteristics: accurate timing of energy storage charging and discharging, charging during off-peak hours and discharging during peak hours; reasonable control of photovoltaic curtailment rate, avoiding blindly pursuing zero curtailment at the expense of economic efficiency; and appropriate reservation of energy storage SOC, maintaining sufficient discharge capacity before peak load. Common problems in poor cases include: premature depletion of energy storage leading to insufficient support during peak load; excessive pursuit of photovoltaic absorption resulting in frequent shallow charging and discharging of energy storage, increasing losses; and lagging energy storage charging and discharging decisions failing to respond promptly to changes in photovoltaic power and load. These regularities are refined into empirical rules for control strategies and integrated into the optimization scheduling algorithm to guide the setting and adjustment of control parameters, thus gaining a regular understanding of strategy optimization.

[0072] Step 6.6, Reinforcement Learning Online Optimization; Based on the continuous accumulation of the running case library output in step 6.1 and the case quality labels output in step 6.4, a reinforcement learning algorithm is used to optimize the control strategy online, resulting in an intelligent control strategy that learns and evolves autonomously. Reinforcement learning is performed based on the continuous accumulation of the running case library and feedback on the running performance. Specifically, the energy management problem is modeled as a Markov decision process. The state space includes system running state characteristics, the action space includes control strategy parameters, and the reward function is defined according to the running performance index; higher performance results in higher rewards, and lower performance results in lower rewards. A deep Q-network algorithm is used for policy learning. The network structure includes an input layer that receives a 64-dimensional state feature vector, a first hidden layer (a fully connected layer with 256 neurons using the ReLU activation function), a second hidden layer (a fully connected layer with 128 neurons), a third hidden layer (a fully connected layer with 64 neurons), and an output layer that outputs the Q-value estimate of the action space dimension. The training process employs an experience replay mechanism and a periodic update strategy for the target network, combined with a greedy exploration strategy to balance exploration and utilization. After each run, the state transition and reward are stored in the experience replay pool, and training samples are randomly sampled from the experience pool to update the neural network parameters. Through extensive trial and error and learning, the algorithm gradually learns to select the optimal control strategy in different scenarios, and the system's autonomous decision-making ability is continuously improved. The trained neural network model is periodically deployed into the system to replace the original manual parameter setting, enabling the automatic evolution and continuous optimization of the control strategy, resulting in an intelligent control strategy that learns and evolves autonomously.

[0073] Step 6.7, A / B testing strategy verification; Based on the performance of the autonomously learning and evolving intelligent control strategy output in step 6.6 during actual operation, A / B testing is used to verify the effectiveness of the new strategy, providing a basis for strategy updates. Specifically, during system operation, the newly optimized control strategy is used in randomly selected time periods, while the original strategy is used in other time periods. The operational effects of the two strategies are compared. Statistical analysis is performed on the improvement of the new strategy relative to the original strategy in various indicators, and hypothesis testing is used to determine whether the improvement is statistically significant. When the new strategy outperforms the original strategy in multiple indicators and does not significantly deteriorate in other indicators, the new strategy is confirmed as effective and officially applied to all operating periods. When the new strategy has no significant effect or negative impact, the original strategy is retained, and the new strategy is further optimized or the optimization direction is abandoned. Through the A / B testing mechanism, it is ensured that each strategy update is based on verified effectiveness, avoiding system performance degradation caused by blind updates, guaranteeing system stability and continuous improvement capabilities, and providing a basis for strategy updates.

[0074] Step 6.8, Cost-Benefit Analysis; Based on the economic data records of the system operation, including the power consumption statistics of each energy path output in step 5.8, such as grid electricity purchase costs, photovoltaic grid connection revenue, and charging service revenue, a cost-benefit analysis method is used to obtain an evaluation of the system's economic operation performance. The cost-benefit analysis is conducted based on the economic data records of the system operation. Specifically, daily operating costs are calculated, including grid electricity purchase expenditure minus photovoltaic grid connection revenue. Energy storage loss depreciation costs are calculated by dividing the total investment in the energy storage system by its expected lifespan and then by 365 days. Equipment maintenance costs are calculated as a fixed percentage of the total equipment investment. Charging service revenue is calculated as the sum of the charging volume multiplied by the charging service price for each time period. Daily net revenue is calculated as charging service revenue minus daily operating costs. Monthly and annual cumulative net revenue is calculated to assess the system's profitability. The investment payback period is calculated as the total system investment divided by the average annual net revenue. Economic indicators are benchmarked against similar projects to identify the advantages and disadvantages of economic performance, thus obtaining an evaluation of the system's economic operation performance.

[0075] Step 6.9: Generate periodic reports; Based on the operational performance evaluation indicators output in step 5.9 and the performance evolution patterns and potential problem alerts output in step 5.10, a periodic report generation mechanism is adopted to obtain operational report documents for different users. Based on the system operational evaluation report, periodic reports are generated. Specifically, monthly operational reports are generated for operations management personnel, including key performance indicators, economic benefit data, equipment health status, and abnormal event records, providing data support for operational decision-making. Equipment status reports are generated for maintenance personnel, including energy storage health scores, equipment operating hours, and maintenance reminders, guiding equipment maintenance work. Economic benefit reports are generated for investors, including detailed income and expenditure statements, investment recovery progress, and comparisons with expected targets, showcasing the project's economic value. All reports are automatically generated using standardized templates and pushed to relevant personnel via email or the system platform, ensuring timely information delivery and obtaining operational report documents for different users.

[0076] In some embodiments, since different users have different preferences for system operation objectives—some users are more concerned with economic efficiency, while others are more concerned with environmental protection—a multi-strategy parallel learning method can be adopted to train optimization strategies for different objective preferences and establish a strategy library. The aim is to provide personalized energy management solutions for users with different needs, enhancing the system's adaptability and flexibility. Specifically, three typical strategy types are defined: economic priority strategy, environmental priority strategy, and equilibrium strategy. The reward function of the economic priority strategy takes minimizing operating costs as the primary objective, with a weight of 0.7, and photovoltaic consumption and equipment lifespan as secondary objectives, each with a weight of 0.15. The reward function of the environmental priority strategy takes maximizing photovoltaic consumption rate as the primary objective, with a weight of 0.7, and economic efficiency and equipment lifespan as secondary objectives. The equilibrium strategy assigns equal weights to the three objectives. DQN models for the three strategies are trained separately to form a strategy library. Users can choose the strategy type to apply according to their own needs, and the system loads the corresponding control model based on the selection. Users can also dynamically switch strategies according to the needs of different time periods; for example, the economic priority strategy is used on days with large peak-valley differences in electricity prices, and the environmental priority strategy is used on environmental protection publicity days. The multi-strategy mechanism meets diverse user needs and enhances the system's market competitiveness.

[0077] In one embodiment of the present invention, a specific example is provided: This invention focuses on the application of integrated photovoltaic, energy storage and charging energy stations in highway service areas. To verify the actual operation effect of the system, a highway service area is selected as an application example. The service area is equipped with a 500 kW photovoltaic system, a 600 kWh energy storage system and 10 120 kW charging piles, with a grid connection capacity of 400 kW.

[0078] A typical operating scenario was selected from the three consecutive days of the National Day holiday in a certain year. During this period, the charging load was higher than usual, posing a severe challenge to the system's scheduling capabilities. The system began implementing a pre-scheduling strategy the day before the holiday. Based on historical data, it predicted a significant increase in the charging load during the holiday and adjusted the energy storage charging plan in advance. During the low electricity price period at night, the energy storage SOC was charged to 85% to prepare for the high load during the holiday.

[0079] The first day of the holiday was sunny and clear, providing favorable conditions for photovoltaic power generation. Typical system operation data for the first day of the holiday is shown in Table 1. Table 1: System operation data during typical periods on the first day of the holiday;

[0080] As shown in Table 1, during peak load periods, the system ensured the normal provision of charging services through the coordinated power supply of photovoltaics, energy storage, and the power grid. The energy storage system continuously discharged throughout the day, effectively alleviating the power supply pressure on the power grid. The power purchased by the grid was always kept within the access capacity range, and no overload occurred.

[0081] The second day was the second day of the holiday, and the morning was cloudy, resulting in significant fluctuations in photovoltaic output. Table 2 shows the photovoltaic output fluctuations and energy storage response data under cloudy weather conditions. Table 2: Data on photovoltaic fluctuations and energy storage response under cloudy weather conditions;

[0082] As can be seen from Table 2, in the face of rapid fluctuations in photovoltaic output, the energy storage system can respond quickly to compensate for power fluctuations and control the power fluctuations transmitted to the grid within a small range, demonstrating a good power smoothing effect.

[0083] Statistics from the three-day holiday show that the system's total photovoltaic power generation was 3600 kWh, with a self-consumption rate of 92%. The charging piles supplied a total of 5800 kWh, of which photovoltaic and energy storage accounted for 68%, and electricity purchased from the grid accounted for 32%. Through the peak-shaving and valley-filling effect of energy storage, the system reduced the demand for electricity from the grid during peak hours, effectively avoiding charging service interruptions caused by insufficient grid capacity. The energy storage system completed 0.8 cycles over the three days, with the charge and discharge depth controlled within a reasonable range, and its health score remained excellent. The overall operating cost over the three days was lower than the control scheme without this invention, the charging service quality was guaranteed, all charging requests were met, the system operated safely and stably, and no malfunctions occurred.

[0084] As can be seen from the above application examples, the energy management system of the photovoltaic-storage-charging integrated device of the present invention can achieve efficient coordinated operation of photovoltaic, energy storage, charging piles and power grid under high load scenarios, effectively cope with charging load fluctuations and photovoltaic output uncertainties, ensure charging service quality, reduce operating costs, and extend energy storage life, and has good practical value and promotion prospects.

[0085] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An integrated photovoltaic, energy storage, and charging energy management system, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous raw data streams, perform preprocessing and state feature extraction, and obtain standard operating state feature vectors. The prediction module, based on the feature vector of standard operating state, performs intelligent prediction of photovoltaic output and charging load, and obtains intelligent prediction results; The optimization scheduling module performs multi-objective energy optimization scheduling based on intelligent prediction results and generates real-time scheduling instructions; The real-time control module performs equipment power control and coordination operations based on real-time scheduling instructions, and obtains equipment power control results. The health management module performs energy storage health status assessment and system performance analysis based on equipment power control results and standard operating status feature vectors, and obtains energy storage health status assessment results and health management results. The strategy learning module builds a case library based on health management results, acquires the characteristics of operating scenarios, learns and optimizes control strategies, verifies the operating effects, and obtains intelligent control strategies that evolve autonomously through learning.

2. The energy management system for the integrated photovoltaic, energy storage, and charging device according to claim 1, characterized in that, The data acquisition module includes: Deploy a distributed sensor network, using industrial Ethernet and ModbusTCP communication protocols to synchronously collect electrical and environmental parameters of measurement points at a preset sampling period, and obtain multi-source heterogeneous raw data streams; Time alignment, data normalization, and wavelet transform denoising are performed on the multi-source heterogeneous raw data stream to obtain the denoised data sequence; Based on the denoised data sequence, the extended Kalman filter algorithm is used to estimate the state of charge and obtain the corrected energy storage SOC value. Based on the denoised data sequence and the corrected energy storage SOC value, principal component analysis is used to extract state features and obtain the standard operating state feature vector.

3. The energy management system for the integrated photovoltaic-storage-charging device according to claim 1, characterized in that, The prediction module includes: A photovoltaic power output prediction model is constructed. Historical operating data of the photovoltaic-storage-charging system is collected as the source domain dataset. A long short-term memory network is used to construct the basic photovoltaic power output prediction model. Data collected from the local photovoltaic-storage-charging system is used as the target domain data to obtain the photovoltaic power output prediction model. A multi-step rolling prediction method is used to generate photovoltaic power prediction values, resulting in photovoltaic output prediction curves at multiple time scales.

4. The energy management system for the integrated photovoltaic, energy storage, and charging device according to claim 1, characterized in that, The prediction module also includes: The quantile regression method is used to estimate the prediction uncertainty. In the output layer of the basic photovoltaic power output prediction model, the quantile loss function is used instead of the traditional mean square error loss function. Multiple prediction models with preset quantiles are trained and correspond to the lower limit, median and upper limit of prediction, respectively, to obtain the photovoltaic power output prediction interval. Based on historical charging records, statistical analysis was conducted. A vehicle arrival model was established based on the Poisson process assumption. The probability distribution of charging power and charging time was fitted using the Gaussian kernel density estimation method. The load was aggregated using the Monte Carlo random sampling method to obtain the charging load probability density distribution.

5. The energy management system for the integrated photovoltaic-storage-charging device according to claim 1, characterized in that, The optimized scheduling module includes: A multi-objective optimization model is constructed, which includes economic objectives, environmental objectives, and equipment life objectives. The economic objective function is to minimize the system's daily operating cost, the environmental objective function is to maximize the photovoltaic absorption rate, and the equipment life objective function is to minimize the equivalent cycle number of the energy storage system. The economic objective function, environmental objective function, and equipment lifespan objective function are transformed to a unified dimension by a normalization method. An adaptive weighting coefficient is introduced to construct a comprehensive objective function in the form of a weighted sum. The weighting coefficient is dynamically adjusted according to the current energy storage health status and electricity price level to obtain a single-objective optimization problem.

6. The energy management system for the integrated photovoltaic-storage-charging device according to claim 1, characterized in that, The optimized scheduling module also includes: Establish a set of constraints for day-ahead energy dispatch, including power balance constraints, energy storage SOC constraints, energy storage charging and discharging power constraints, grid interaction power constraints, charging pile output power constraints, and energy storage charging and discharging state mutual exclusion constraints, to obtain a complete day-ahead optimal dispatch mathematical model; The optimal scheduling plan for energy storage charging and discharging power, grid interaction power, and photovoltaic utilization power in each time period of the future preset time period is obtained by using a mixed integer linear programming algorithm. A rolling optimization strategy is adopted to establish a real-time energy scheduling model. A preset time period is used as a time period, and the optimization is performed once every preset time interval to obtain a refined power allocation instruction. The power allocation instruction of the current time period is extracted to obtain the real-time scheduling instruction to be sent to the underlying controller.

7. The energy management system for the integrated photovoltaic-storage-charging device according to claim 1, characterized in that, The real-time control module includes: The power command is sent to the photovoltaic inverter, energy storage converter and charging pile controller using a standardized industrial communication protocol, and feedback signals confirming receipt are obtained from each device. A power regulation algorithm based on droop control is used to control the charging and discharging behavior of the energy storage system. The energy storage converter adopts a dual closed-loop control structure, with the outer loop being the power control loop and the inner loop being the current control loop. When multiple energy storage converters are connected in parallel, the droop control strategy is used to achieve automatic power distribution and obtain the DC bus voltage and actual charging and discharging power.

8. The energy management system for the integrated photovoltaic, energy storage, and charging device according to claim 1, characterized in that, The health management module includes: The charge-discharge cycles of energy storage SOC are counted using the rainflow counting method. The peak and valley values ​​in the SOC sequence are identified. Peak-valley pairs are matched to form cycles according to the rainflow rule. Each cycle is characterized by cycle depth and average SOC, and the cycle number distribution is obtained. The equivalent number of full cycles of the energy storage system is calculated using the linear cumulative damage theory. Based on the cycle life corresponding to different cycle depths, the lifetime damage caused by each actual cycle is equivalent to the number of full cycles, thus obtaining an estimated value of energy storage capacity decay. A comprehensive scoring method is used to calculate the energy storage health score. Based on the score, the energy storage health status is divided into four levels: excellent, good, average, and declining, thus obtaining the energy storage health status index.

9. The energy management system for the integrated photovoltaic-storage-charging device according to claim 1, characterized in that, The health management module also includes: An isolated forest anomaly detection algorithm is used to identify abnormal operating modes of energy storage systems. Multiple isolated trees are constructed to calculate the average path length of data points. When the anomaly score of real-time monitoring data exceeds a preset threshold, an anomaly warning is triggered to obtain a potential fault warning signal. Based on the energy storage health status indicators, an adaptive strategy is used to dynamically correct the operating constraint parameters of the energy storage system. Different SOC operating ranges and charging / discharging power limits are set for energy storage units with different health levels to obtain the energy storage operation strategy.

10. The energy management system for the integrated photovoltaic, energy storage, and charging device according to claim 1, characterized in that, The strategy learning module includes: Each system run is recorded as a case, and each case includes the characteristics of the running scenario, control strategy parameters, and running effect indicators, thus obtaining a running case library; Define a similarity measurement function for scene features, calculate the comprehensive similarity between the current scene and historical cases, and select the top few cases with the highest similarity as a set of reference cases; The energy management problem is modeled as a Markov decision process, and a deep Q-network algorithm is used for policy learning. State transitions and rewards are stored in an experience replay pool, and the neural network parameters are updated by randomly sampling training samples to obtain an intelligent control strategy that learns and evolves autonomously.

Images