Adaptive energy scheduling method and system for distributed photovoltaic energy storage system

Abstract

The application discloses a kind of distributed photovoltaic energy storage systems adaptive energy scheduling method, comprising: obtaining the multi-source data of distributed photovoltaic energy storage system, the multi-source data is preprocessed;The preprocessed multi-source data is input into LSTM neural network model, and output prediction result;Build multi-objective optimization model containing minimization power consumption cost and maximization local consumption rate;The improved particle swarm algorithm is used to solve the multi-objective optimization model, and the optimal charge-discharge power instruction sequence is output;On this basis, calculate rolling prediction error;When rolling prediction error exceeds preset error threshold, trigger feedback correction;Cloud platform regularly based on historical deviation data parameter fine tuning is carried out to LSTM prediction model, and according to different seasons and weather mode adaptive update strategy parameter of improved particle swarm algorithm, form closed loop self-learning mechanism.The application realizes adaptive energy scheduling to distributed photovoltaic energy storage system under complex working condition.

Description

Technical Field

[0001] This invention relates to the field of microgrid operation control and integrated energy management technology, specifically a distributed photovoltaic-storage adaptive energy dispatching method and system for industrial and commercial parks and residential communities. Background Technology

[0002] In recent years, the large-scale application of distributed photovoltaic energy storage systems in industrial parks has accelerated, becoming a core technological solution for reducing electricity costs, improving energy reliability, and promoting green and low-carbon transformation in these parks. However, in actual project operation, existing energy dispatch strategies generally suffer from static rigidity and insufficient adaptive capabilities, making it difficult to match the complex and ever-changing energy consumption and power generation conditions in industrial parks. Specific shortcomings are as follows: First, power forecasting relies on empirical rules and has poor adaptability to extreme weather conditions. Currently, most dispatching systems still use fixed-period charging and discharging strategies or rely solely on simple historical averages to forecast photovoltaic output and load, without using intelligent models to adapt to weather fluctuations. Taking Hainan as an example, sudden weather events such as a sharp drop in photovoltaic output caused by afternoon convective clouds occur frequently. Traditional static forecasting cannot quickly track sudden changes in output, which can easily lead to a serious deviation between the dispatching plan and the actual operating conditions. Energy storage is forced to passively replenish energy, which not only significantly reduces the economic efficiency of system operation but also adds an extra burden to battery charging and discharging.

[0003] Second, the optimization scheduling strategy is "one-size-fits-all" and lacks multi-objective coordination capabilities. Although some solutions use optimization algorithms, they fail to incorporate dynamic coupling factors such as battery life loss, dynamic charging and discharging efficiency, time-of-use pricing, and real-time fluctuations in park load into the constraints. The scheduling strategy is simple and rigid, and cannot simultaneously take into account the dual objectives of "optimal electricity cost" and "protection of energy storage battery life". It is easy to have the short-sighted problem of unilaterally pursuing economic efficiency at the expense of battery damage.

[0004] Third, the scheduling system lacks closed-loop iteration capability, leading to a decline in adaptability over long-term operation. Existing scheduling schemes generally lack real-time feedback and correction mechanisms for prediction errors and execution deviations. After scheduling instructions are issued, actual operating data cannot be fed back to optimize the model, resulting in a continuous disconnect between the strategy and on-site conditions after long-term operation. At the same time, it relies heavily on manual periodic parameter tuning, resulting in high operation and maintenance costs and slow response speed, making it difficult to support the large-scale promotion of photovoltaic and energy storage systems in parks.

[0005] Therefore, there is an urgent need for an adaptive energy dispatching method and system for distributed photovoltaic energy storage systems to solve the above problems. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the aforementioned technologies. Therefore, a first aspect of the present invention aims to propose an adaptive energy scheduling method for distributed photovoltaic energy storage systems to address the problem that edge computing devices in industrial parks have limited computing power, making it impossible to run heavy-duty prediction and optimization models, while cloud platforms struggle to obtain real-time operational data for rapid iteration.

[0007] The second objective of this invention is to propose an adaptive energy dispatching system for a distributed photovoltaic energy storage system.

[0008] To achieve the above objectives, a first aspect of the present invention proposes an adaptive energy dispatch method for a distributed photovoltaic energy storage system, comprising: Meteorological data from distributed photovoltaic energy storage systems are collected using meteorological sensors. Combined with historical load curves and historical photovoltaic power generation obtained from user-side smart meters, multi-source data is constructed and preprocessed. The preprocessed multi-source data is input into the LSTM neural network model, and the prediction results are output. A rolling time-domain update mechanism is adopted to periodically update the model input time series with the latest measured data, recalculate and output the updated prediction results, and store them in the local database. The prediction results include the photovoltaic power generation prediction curve and the user load prediction curve for a preset period of time in the future. Using the peak-valley flat electricity price of the power grid, user load forecasting, and the state of charge (SOC) of the energy storage battery as boundary conditions, a multi-objective optimization model is constructed, which includes minimizing electricity costs and maximizing local absorption rate. An improved particle swarm optimization algorithm is used to solve the multi-objective optimization model and output the optimal charging and discharging power command sequence of the energy storage system within a preset time period in the future. The dispatch controller periodically compares the predicted photovoltaic power generation, the predicted user load, and the corresponding measured values ​​to calculate the rolling prediction error. When the rolling prediction error exceeds the preset error threshold, feedback correction is triggered, and the energy storage charging and discharging power limits and battery state of charge reference trajectory are dynamically adjusted for subsequent dispatch periods. The scheduling execution data, prediction error distribution, and optimization target achievement status are uploaded to the cloud platform. The cloud platform regularly fine-tunes the parameters of the LSTM prediction model based on historical deviation data, and adaptively updates and improves the strategy parameters of the particle swarm algorithm according to different seasons and weather patterns, forming a closed-loop self-learning mechanism.

[0009] Preferably, the multi-source data is preprocessed, including: Linear interpolation was used to align and correct the timestamps collected from different devices. The 3σ criterion combined with the Laida criterion is used to detect outliers in time series data, and to remove abnormal and invalid data caused by sudden changes in irradiance, load spikes, zero-value dead zones of photovoltaic power, and sensor fault drift. A scenario-based completion strategy is adopted to complete missing data. The preprocessed meteorological data, historical load sequences, and historical photovoltaic power sequences are reconstructed into a fixed-length model input time sample sequence in chronological order.

[0010] Preferably, the LSTM prediction model includes an input layer, a hidden layer, and an output layer; The number of neurons in the input layer is consistent with the feature dimension of the preprocessed data. Two hidden layers are set, with 64 neurons in each layer, and the ReLU activation function is used; The output layer has two neurons, corresponding to the predicted photovoltaic power generation and the predicted user electricity load for a preset time period, respectively. The output layer uses a linear activation function.

[0011] Preferably, a rolling time-domain update mechanism is adopted, which periodically updates the model input time series with the latest measured data and recalculates and outputs the updated prediction results, including: An update cycle consistent with the sampling period is adopted. The model input time series is updated periodically using the latest measured data. The oldest sampling point in the input series is removed, the latest measured sampling point is added, and the data is re-input into the LSTM prediction model for calculation, and the updated prediction result is output.

[0012] Preferably, an improved particle swarm optimization algorithm is used to solve the multi-objective optimization model, outputting the optimal charging and discharging power command sequence of the energy storage system within a preset time period, including: S311: Initialize the particle swarm; S312: For the charge and discharge power sequence corresponding to each initial particle, perform two nonlinear constraints on battery charge and discharge efficiency and lifetime loss, calculate the actual charge and discharge power and calculate the loss value through the lifetime loss function, remove particles that do not meet the constraints, and reinitialize the particles. S313: Based on the comprehensive objective function, a lifetime loss penalty term is added to calculate the fitness value of each particle; S314: Based on the convergence judgment factor, determine the current state of the algorithm, adjust the inertia weight and learning factor, and update the particle velocity and position; S315: Determine if the iteration termination condition has been met: If it has, output the optimal position of the group, i.e. the optimal charge / discharge power command sequence; if it has not been met, return to step S312 to continue iterating until the termination condition is met.

[0013] Preferably, the dispatch controller periodically compares the predicted photovoltaic power generation, the predicted user load, and the corresponding measured values ​​to calculate the rolling prediction error, including: The dispatch controller synchronously collects the measured values ​​of photovoltaic power and user load according to a preset cycle, and calls the corresponding photovoltaic power prediction values ​​and load prediction values ​​in the local database to calculate the rolling prediction error; the rolling prediction error calculation adopts a dual judgment of absolute error and relative error.

[0014] Preferably, after the correction mechanism is triggered, the scheduling controller automatically adjusts the energy storage charging and discharging power limits and the battery state-of-charge reference trajectory for subsequent scheduling periods. Specific adjustment strategies include: If the predicted photovoltaic power value is consistently lower than the measured value, the limit for energy storage charging power will be increased while the limit for discharging power will be decreased to increase local photovoltaic consumption; if the predicted photovoltaic power value is consistently higher than the measured value, the limit for charging power will be decreased while the limit for discharging power will be increased. If the load forecast value is consistently higher than the measured value, the SOC reference trajectory for subsequent periods should be appropriately reduced to decrease the energy storage charging amount; if the load forecast value is consistently lower than the measured value, the SOC reference trajectory should be appropriately increased to increase the energy storage charging amount.

[0015] Preferably, the cloud platform periodically fine-tunes the parameters of the LSTM prediction model based on historical bias data, including: The cloud platform regularly summarizes and analyzes the uploaded historical data, focusing on extracting historical prediction deviation data, and uses the gradient descent method to fine-tune the parameters of the LSTM prediction model.

[0016] Preferred closed-loop self-learning mechanisms include: The cloud platform sends the updated algorithm strategy parameters to the local scheduling controller. The local controller combines the updated model and algorithm parameters to execute the new prediction, optimization and scheduling process. At the same time, it continuously collects running data and uploads it to the cloud platform, forming a closed-loop self-learning mechanism of local scheduling - data upload - cloud platform optimization - parameter distribution - local update.

[0017] To achieve the above objectives, a second aspect of the present invention provides an adaptive energy dispatch system for a distributed photovoltaic energy storage system, comprising: The data acquisition module is used to collect meteorological data of the distributed photovoltaic energy storage system based on meteorological sensors, combine it with the historical load curve and historical photovoltaic power generation obtained from the user-side smart meter to construct multi-source data, and preprocess the multi-source data. The prediction calculation module is used to input preprocessed multi-source data into an LSTM neural network model and output prediction results; and adopts a rolling time-domain update mechanism to periodically update the model input time series with the latest measured data and recalculate and output updated prediction results; wherein, the prediction results include a photovoltaic power generation prediction curve and a user load prediction curve for a preset period of time in the future; The optimization scheduling module is used to construct a multi-objective optimization model that includes minimizing electricity costs and maximizing local absorption rate, using the grid peak-valley flat electricity price, user load forecast, and energy storage battery state of charge as boundary conditions; the improved particle swarm optimization algorithm is used to solve the multi-objective optimization model and output the optimal charging and discharging power command sequence of the energy storage system within a preset time period in the future; The feedback control module is used by the dispatch controller to periodically compare the photovoltaic power generation forecast, user load forecast and corresponding measured values ​​to calculate the rolling forecast error. When the rolling forecast error exceeds the preset error threshold, feedback correction is triggered and the energy storage charging and discharging power limit and battery state of charge reference trajectory are dynamically adjusted for subsequent dispatch periods. The cloud platform management module is used to upload scheduling execution data, prediction error distribution, and optimization target achievement status to the cloud platform. The cloud platform regularly fine-tunes the parameters of the LSTM prediction model based on historical deviation data, and adaptively updates and improves the strategy parameters of the particle swarm algorithm according to different seasons and weather patterns, forming a closed-loop self-learning mechanism.

[0018] The present invention has achieved the following beneficial effects: This invention employs an LSTM neural network combined with a rolling time-domain update mechanism to accurately track photovoltaic power fluctuations caused by sudden weather changes, significantly improving the accuracy of short-term power and load forecasting, demonstrating significant advantages over traditional methods. By incorporating an improved particle swarm optimization algorithm with embedded engineering constraints to formulate scheduling strategies, it balances electricity economy with the lifespan of energy storage batteries, avoiding short-sighted scheduling behavior that damages batteries in pursuit of cost savings. Relying on a real-time prediction error feedback correction mechanism coupled with dynamic SOC safety management, it effectively suppresses the accumulation of prediction errors, significantly reducing the risk of battery overcharging and over-discharging, and ensuring stable and safe system operation. The cloud platform's closed-loop self-learning mechanism enables the system to have continuous optimization and adaptive capabilities, achieving increasingly accurate scheduling with use, greatly reducing the frequency of manual parameter tuning, and making it more suitable for the large-scale deployment and long-term operation and maintenance of distributed photovoltaic energy storage systems.

[0019] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a flowchart of an adaptive energy dispatching method for a distributed photovoltaic energy storage system according to an embodiment of the present invention; Figure 2 This is a flowchart of preprocessing the multi-source data according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the solution of the multi-objective optimization model using an improved particle swarm optimization algorithm according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating how a scheduling controller automatically adjusts the energy storage charging and discharging power limits and the battery state-of-charge reference trajectory for subsequent scheduling periods, according to an embodiment of the present invention. Figure 5 This is a block diagram of an adaptive energy dispatch system for a distributed photovoltaic energy storage system according to an embodiment of the present invention. Detailed Implementation

[0022] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0023] This application provides an adaptive energy dispatching method and system for distributed photovoltaic energy storage systems. By constructing a closed-loop adaptive energy dispatching system that integrates prediction, optimization, correction, and self-learning, it improves the accuracy of photovoltaic and load forecasting from the source, dynamically optimizes dispatching strategies, corrects operational deviations in real time, and continuously iterates model parameters. It can accurately adapt to complex operating conditions such as sudden weather changes, while simultaneously taking into account operational economy, battery life, and system safety. It can also rely on cloud platform self-learning to achieve continuous strategy optimization, significantly reducing manual operation and maintenance costs, and perfectly adapting to the large-scale deployment and long-term stable and efficient operation of distributed photovoltaic energy storage systems in parks.

[0024] like Figure 1 As shown in the figure, this embodiment discloses an adaptive energy dispatch method for a distributed photovoltaic energy storage system, which specifically includes the following steps: Step S1: Collect meteorological data of the distributed photovoltaic energy storage system based on meteorological sensors, combine it with the historical load curve and historical photovoltaic power generation obtained from the user-side smart meter, construct multi-source data, and preprocess the multi-source data.

[0025] Specifically, multiple types of meteorological sensors are rationally deployed at the distributed photovoltaic power station site, including: irradiance sensors, ambient temperature sensors, photovoltaic module temperature sensors, wind speed sensors, and humidity sensors. All meteorological sensors adopt the RS485 communication protocol and synchronously collect raw meteorological data from multiple sources at a preset sampling period of 5 minutes. The sampling period is consistent with the subsequent rolling time-domain prediction period to ensure data synchronization in the time dimension. Simultaneously, a local industrial communication gateway (supporting Modbus-RTU / TCP protocols) connects to the user-side smart meters to read historical moment-by-moment user electricity load curves stored in the smart meters in real time, as well as historical photovoltaic power generation time-series data uploaded by the photovoltaic inverters and statistically analyzed by the smart meters. The historical load curves are selected from the intraday time-of-use active power time-series sequences of the past 7 and 30 days, and the historical photovoltaic power generation is selected from the photovoltaic output time-series data of the corresponding period. The timestamps of all data are uniformly calibrated to Beijing time, completing the aggregation of multi-source heterogeneous data (meteorological data, historical load data, and historical photovoltaic power data) and storing it in the database of the local edge computing gateway.

[0026] like Figure 2 As shown, after the multi-source data aggregation is completed, preprocessing operations are performed on the multi-source data in sequence. The specific steps are as follows: Since there may be slight deviations in the clocks of meteorological sensors, smart meters, and photovoltaic inverters, which may cause time misalignment in the sampling data, a linear interpolation method is used to correct the timestamps of different devices. The original data with different sampling frequencies are uniformly resampled into time series data with a fixed time interval of 5 minutes to ensure that all data correspond one-to-one in the time dimension and eliminate the impact of time misalignment on subsequent model predictions.

[0027] The 3σ criterion combined with the Laida criterion is used to detect outliers in time series data. Three times the standard deviation of the data is set as the outlier threshold. Abnormal and invalid data such as sudden changes in irradiance, load spikes, zero-value dead zones of photovoltaic power (photovoltaic output continuously at 0 for more than 30 minutes during the day without any shading records) and sensor fault drift are removed. At the same time, obviously erroneous data such as photovoltaic output greater than 0 at night and negative load power are directly marked and removed to ensure the rationality of the data.

[0028] To address data gaps caused by sensor offlineness or communication interruptions, a scenario-based completion strategy is adopted: for short-term continuous data gaps, linear interpolation of data from adjacent time periods is used for completion; for long-term continuous data gaps, the mean and trend of similar historical time periods are combined for completion to ensure the continuity and completeness of time-series data.

[0029] Because the dimensions and numerical ranges of multi-source data such as irradiance, temperature, wind speed, load power, and photovoltaic power differ significantly (e.g., irradiance is measured in W / m², and load power in kW), they can affect the training convergence speed and prediction accuracy of the subsequent LSTM neural network model. Therefore, a maximum-minimum normalization method is adopted to map all data to the [0, 1] interval. The normalization formula is as follows: ,in This is the original data. The minimum value of this type of data. The maximum value of this type of data. These are the normalized values, and the normalized data is stored in the model input format.

[0030] The preprocessed meteorological data, historical load sequences, and historical photovoltaic power sequences are reconstructed into a fixed-length model input time sample sequence in chronological order. Each sample contains all multi-source data features within the corresponding time window, completing the entire preprocessing process. The preprocessed data is then directly input into the subsequent LSTM neural network model for prediction.

[0031] Step S2: Input the preprocessed multi-source data into the LSTM neural network model and output the prediction results; and adopt a rolling time-domain update mechanism to periodically update the model input time series with the latest measured data, recalculate and output the updated prediction results, and store them in the local database; wherein, the prediction results include the photovoltaic power generation prediction curve and the user load prediction curve for a preset period of time in the future.

[0032] An LSTM neural network prediction model was built, employing a three-layer structure (input layer, hidden layer, and output layer). Specific parameter settings are as follows: the number of neurons in the input layer matches the dimension of the preprocessed data features (e.g., 5 meteorological features, 1 historical load feature, and 1 historical photovoltaic power feature, totaling 7 features; the number of neurons in the input layer is set to 7); two hidden layers are set, each with 64 neurons, using the ReLU activation function to avoid the vanishing gradient problem, and the dropout coefficient for the hidden layers is set to 0.2 to prevent overfitting; the output layer has 2 neurons, corresponding to the predicted photovoltaic power generation and the predicted user electricity load over a preset time period, respectively, and uses a linear activation function to ensure the continuity of the predicted values.

[0033] Model training process: The preprocessed time series data of the past 3 months are selected as the training set (80%) and the time series data of the past 1 month are selected as the test set (20%). The mean squared error (MSE) is used as the loss function, and the Adam optimizer is used for model training. The learning rate is set to 0.001 and the number of training iterations is set to 100 rounds. When the loss function of the test set converges (the change in the loss value is less than 0.0001 for 10 consecutive rounds), the training is stopped, and the trained LSTM prediction model is obtained and stored in the local edge computing gateway for real-time prediction.

[0034] Prediction Execution Process: The preprocessed latest time sample sequence is input into the trained LSTM prediction model, which outputs prediction results for a preset future duration, preferably 24 hours (short-term intraday prediction). The prediction time interval is consistent with the sampling period (5 minutes), resulting in 48 prediction points. This forms a complete photovoltaic power generation prediction curve and a user electricity load prediction curve. The prediction results are stored in the local database and simultaneously synchronized to the dispatch controller. It is important to note that the rolling time-domain update mechanism in step S2 only applies to the input data layer of the LSTM prediction model. Its purpose is to correct the future prediction curve trend by absorbing the latest measured information. This process belongs to the micro-level data perception and prediction correction and does not change or interfere with the physical safety boundary and strategy optimization boundary of the energy storage system in step S3, thereby ensuring the logical decoupling between the prediction layer and the control layer.

[0035] Furthermore, a rolling time-domain update mechanism is implemented: An update cycle consistent with the sampling period is adopted. The latest measured data (including the latest collected meteorological data, the latest measured photovoltaic power, and the latest measured user load) is periodically used to update the model's input time series. The oldest sampling point in the input series is removed, and the latest measured sampling point is added. The data is then re-input into the LSTM prediction model for computation, outputting an updated 24-hour forecast curve. This rolling update method enables dynamic adjustment of the prediction results, promptly correcting prediction deviations caused by changes in operating conditions (such as sudden weather changes or load fluctuations), and ensuring that the prediction accuracy meets scheduling requirements.

[0036] Step S3: Using the grid peak-valley flat electricity price, user load forecast, and energy storage battery state of charge (SOC) as boundary conditions, construct a multi-objective optimization model that includes minimizing electricity costs and maximizing local absorption rate; use an improved particle swarm optimization algorithm to solve the multi-objective optimization model and output the optimal charging and discharging power command sequence of the energy storage system within a preset time period in the future.

[0037] First, the boundary constraints of the optimization model are clearly defined. Combining the grid peak-valley flat electricity price, user load forecasting, photovoltaic power forecasting, and the state of charge (SOC) of the energy storage battery, the specific constraints are set as follows: Electricity Price Constraint: The local grid peak-valley flat time-of-use electricity price is adopted. Electricity price data is obtained in real-time from the grid marketing system via a communication module to ensure the timeliness of electricity price information. Load Constraint: Real-time user-side electricity load must be prioritized, i.e., photovoltaic output + energy storage discharge power ≥ real-time user load. When photovoltaic output + energy storage discharge power is insufficient, the grid will supplement the power supply; when photovoltaic output + energy storage discharge power is excessive, the excess electricity can be used for energy storage charging or grid connection. Photovoltaic Power Constraint: The predicted photovoltaic power generation value serves as the upper limit of the constraint. The actual photovoltaic output must not exceed 1.1 times the predicted value and must not be lower than 0.9 times the predicted value to avoid dispatch disorder caused by sudden changes in photovoltaic output. Energy storage battery constraints: The state of charge (SOC) of the energy storage battery is collected in real time (via the battery management system BMS, with a collection cycle of 5 minutes). The SOC constraint range is 20%-80% to avoid overcharging and over-discharging and extend battery life. Energy storage charging and discharging power constraints: The upper limit of charging power is 1.0 times the rated power of the energy storage, and the upper limit of discharging power is 1.0 times the rated power of the energy storage. This constraint is an insurmountable hardware safety limit. In the subsequent optimization solution and feedback correction stages, all calculated charging and discharging power commands must undergo saturation limiting processing before being sent to the PCS (energy storage converter) to ensure that the actual issued commands do not exceed the physical rated capacity of the equipment, fundamentally eliminating the risk of hardware overload damage. Charging and discharging power must not be negative (power is positive during charging and negative during discharging). Secondly, a multi-objective optimization scheduling model is constructed to minimize the user-side electricity cost and maximize the local photovoltaic absorption rate as dual optimization objectives. The specific objective function is set as follows: Objective function 1 (minimize electricity cost): ,in, The total daily electricity cost for the user (in yuan). This represents the number of sampling points per day (48, 5 minutes / point). Let t be the power supplied by the grid at time t (kW). The time-of-use electricity price at time t (yuan / kWh); For time step, Objective function 2 (maximizing local solar grid integration rate): ,in, The local photovoltaic grid connection rate (%) Let t be the photovoltaic power generation (kW). Let t be the grid-connected photovoltaic power (kW). When the value is 0, the local absorption rate is the highest. This represents the number of sampling points per day.

[0038] To balance the two optimization objectives, a weighted summation method is used to transform the multi-objective optimization problem into a single-objective optimization problem. The weight coefficients are set according to user needs (electricity cost weight 0.6, local absorption rate weight 0.4, which can be flexibly adjusted), resulting in the comprehensive objective function: (Transforming the maximization objective into a minimization objective facilitates the solution, where,) for The value after normalization; for The value after converting to a decimal; (For the comprehensive objective function).

[0039] Then, an improved particle swarm optimization algorithm is used to solve the multi-objective optimization model. The improvements are mainly reflected in: dynamically adjusting the search strategy to avoid premature convergence, and embedding nonlinear constraints such as battery charging and discharging efficiency and lifespan loss to make the solution results closer to engineering practice. The specific improvement measures and solution steps are as follows: The specific improvements are as follows: These improvements address the issues of traditional particle swarm optimization algorithms, such as getting trapped in local optima and slow convergence, while also considering practical engineering needs. To effectively avoid premature convergence in the algorithm, the search strategy is dynamically and adaptively adjusted. Specifically, the initial inertia weight is set to 0.9, which linearly decays to 0.4 as the iteration progresses, while a convergence judgment factor is added. When convergence is detected, the system automatically switches to a local fine-grained search mode, improving local search accuracy by reducing the decay rate of the inertia weight and increasing the proportion of the local learning factor. If the algorithm gets stuck in a local optimum, it switches to a global wide-area search mode, expanding the search range by increasing the inertia weight and mutation probability, thus achieving a dynamic balance between global and local search capabilities and fundamentally avoiding the risk of premature convergence.

[0040] The learning factors of the algorithm are dynamically optimized to improve convergence performance. This specifically includes optimizing individual cognitive factors. The group social factor decreased linearly from 2.5 to 1.0. The value increases linearly from 1.0 to 2.5. This design effectively enhances the convergence accuracy of the algorithm in the later stages, and at the same time, it forms a synergistic effect with the aforementioned dynamic search strategy switching, realizing the dynamic adaptation of the algorithm's search capability and convergence requirements, and further optimizing the solution effect.

[0041] To ensure the practicality of the solution results, two major nonlinear constraints are embedded, tailored to actual engineering applications: one is the charging and discharging efficiency constraint, determined by the charging efficiency coefficient. (Value taken as 0.95), discharge efficiency coefficient (Value 0.92) corrects the actual charging and discharging power to accurately match the actual working efficiency characteristics of the energy storage battery; secondly, it imposes a lifespan loss constraint, using a cycle lifespan loss function. (in, The loss coefficient is... This is the change in state of charge. For charging and discharging power, This refers to the charging and discharging time. The battery loss is quantified (to account for cycle life loss), and this loss value is included in the penalty term of the fitness function. This design ensures that the solution meets both economic requirements and battery life considerations, perfectly aligning with actual engineering operation needs.

[0042] By strengthening the mutation operation, the algorithm's global search capability is further enhanced, ensuring the global optimality of the solution: after every 10 iterations, a mutation operation is performed on 10% of the particles (randomly adjusting particle positions); for particles trapped in local optima, a forced mutation measure is taken to increase their mutation probability to 20%. This design can effectively break the local optimum dilemma, further avoid premature convergence problems, and ensure that the algorithm can find the global optimal solution.

[0043] like Figure 3As shown, the solution steps are as follows: Combining the improvement measures, the solution is obtained by following these steps, ensuring that each step reflects the improvement points and aligns with engineering practice; S311: Initialize the particle swarm: The number of particles is set to 50, and the particle dimension is consistent with the prediction duration (48 dimensions, corresponding to the energy storage charging and discharging power every 5 minutes in the next 24 hours). The particle position range corresponds to the energy storage charging and discharging power constraint (-P rated ~ P rated). Simultaneously, parameters such as the convergence judgment factor, efficiency coefficient, and lifetime loss coefficient are initialized to lay the foundation for subsequent dynamic search and nonlinear constraint verification. S312: Nonlinear constraint verification: For the charging and discharging power sequence corresponding to each initial particle, the battery charging and discharging efficiency and lifetime loss are verified as two major nonlinear constraints. The actual charging and discharging power is calculated, and the loss value is calculated using the lifetime loss function. Particles that do not meet the constraints are removed, and these particles are reinitialized to ensure that the charging and discharging strategy corresponding to the particles meets the actual engineering requirements. S313: Calculate the fitness value of each particle: Based on the comprehensive objective function, a lifetime loss penalty term is added (the penalty coefficient is set to 0.1; the larger the loss value, the worse the fitness value). The fitness value of each particle is calculated, taking into account electricity cost, local absorption rate, and battery life, to achieve a balance between multiple objectives and actual engineering needs. S314: Dynamic search strategy switching and particle update (avoiding premature convergence): Based on the convergence judgment factor, the current state of the algorithm (global search / local search) is determined, and the inertia weight and learning factor are adjusted to update the particle velocity and position. Forced mutation is performed on particles trapped in local optima, and local fine-tuning search is performed on particles that meet the convergence conditions to ensure that the algorithm does not converge prematurely and improve the solution accuracy. S315: Determine if the iteration termination condition is met (100 iterations or fitness value convergence): If met, the optimal position of the population is output, i.e., the optimal charging and discharging power command sequence of the energy storage system for the next 24 hours (which has satisfied nonlinear constraints and is close to engineering practice); if not met, return to step S312 to continue iteration until the termination condition is met.

[0044] After the solution is completed, the optimal charge / discharge power command sequence is synchronized to the scheduling controller. This command sequence has fully considered the actual engineering constraints such as battery charge / discharge efficiency and life loss, and ensures that the solution result is globally optimal through a dynamic search strategy. The command sequence includes energy storage charge / discharge power commands every 5 minutes, clearly defining the three working states of charging, discharging, and shutdown, providing a precise and realistic basis for subsequent scheduling execution.

[0045] Step S4: The dispatch controller periodically compares the photovoltaic power generation forecast, the user load forecast, and the corresponding measured values ​​to calculate the rolling forecast error. When the rolling forecast error exceeds the preset error threshold, feedback correction is triggered, and the energy storage charging and discharging power limits and battery state of charge reference trajectory are dynamically adjusted for subsequent dispatch periods.

[0046] Specifically, to avoid system overshoot oscillation caused by the dual regulation of high-frequency rolling prediction in step S2 and feedback correction in step S4, this embodiment constructs a feedback control mechanism that decouples microscopic prediction correction (S2) from macroscopic boundary reconstruction (S4). The scheduling controller calculates the rolling prediction error according to a preset time period, and the specific implementation process is as follows: First, the rolling forecast error is calculated and systematic deviations are identified. The dispatch controller synchronously collects measured photovoltaic power values ​​(from the photovoltaic inverter) and measured user load values, and calls the corresponding photovoltaic power forecast values ​​and load forecast values ​​from the local database. The rolling forecast error calculation uses both absolute and relative error judgments. The specific calculation formula is as follows: The absolute error calculation formula is: The formula for calculating relative error is: ;in, The measured value at time t Let S be the predicted value at time t. Next, set the anti-jitter filter trigger conditions. To filter out the random high-frequency fluctuations already absorbed in step S2, the trigger condition for S4 does not target single accidental errors but focuses on systematic deviations. Based on actual industrial application needs, the relative error threshold for photovoltaic power prediction is set at 15%, and the relative error threshold for user load prediction is set at 12%. When the relative errors of three consecutive sampling points (i.e., a 15-minute window) all exceed the corresponding threshold, or the relative error of a single sampling point exceeds 1.5 times the threshold (i.e., 22.5% for photovoltaic and 18% for load), a systematic failure of the prediction model is determined, triggering the local macroscopic feedback correction mechanism.

[0047] Then, macroscopic boundary reconstruction is performed. After the correction mechanism is triggered, the scheduling controller does not directly modify the real-time power command of the underlying PCS, but dynamically adjusts the underlying constraint boundary parameters of the multi-objective optimization model in step S3, and triggers S3 to re-optimize, such as... Figure 4 As shown, the specific adjustment strategies are as follows: (1) Adjustment of the upper limit of charging and discharging power scheduling: If the photovoltaic power prediction value is continuously lower than the actual value, it indicates that the photovoltaic output is excessive. Under the premise of strictly adhering to the physical hard limit of 1.0 times the rated power of energy storage, the upper limit of charging power scheduling in the optimization model of step S3 is lowered, forcing the optimization algorithm to reduce charging instructions; if the photovoltaic power prediction value is continuously higher than the actual value, it indicates that the photovoltaic output is insufficient, and the upper limit of discharging power scheduling is raised. (2) SOC reference trajectory translation and reconstruction: If the load prediction value is continuously higher than the actual value, it indicates that the load demand is less than expected. The SOC reference trajectory benchmark value of subsequent scheduling periods is lowered as a whole to release the energy storage charging space; if the load prediction value is continuously lower than the actual value, it indicates that the load demand is greater than expected, and the SOC reference trajectory benchmark value is raised to reserve the discharging space.

[0048] Finally, an action lockout mechanism is introduced to prevent control oscillations. To prevent frequent triggering of S4 from causing repeated jumps in the optimization boundary and leading to system oscillations, the cooling lockout period for the S4 correction action is set to 30 minutes (i.e., 6 sampling points). During the 30 minutes following S4 triggering and boundary reconstruction, the system automatically disables the S4 triggering logic, relying solely on the rolling prediction of step S2 and the routine optimization of step S3 for micro-adjustments. The error monitoring and triggering logic of S4 can only be reactivated after the lockout period ends.

[0049] After the correction is completed and the lockout period ends, the scheduling controller reissues the charging and discharging commands according to the adjusted optimized boundary parameters, and records the correction process data (error value, adjustment parameters, adjustment time, lockout status) for subsequent cloud platform model optimization, ensuring the adaptive adjustment capability of the scheduling strategy and the smooth and stable operation of the system.

[0050] Step S5: Upload the scheduling execution data, prediction error distribution, and optimization target achievement status to the cloud platform; the cloud platform regularly fine-tunes the parameters of the LSTM prediction model based on historical deviation data, and adaptively updates and improves the strategy parameters of the particle swarm algorithm according to different seasons and weather patterns, forming a closed-loop self-learning mechanism.

[0051] The cloud platform enables offline iterative optimization of model and algorithm parameters, forming a closed loop between local real-time scheduling and cloud platform offline optimization. This improves the long-term adaptability of the scheduling method. The specific implementation process is as follows: First, the local edge computing gateway uploads local scheduling execution data, prediction error distribution data, and optimization target achievement status to the cloud platform via a 4G / 5G communication module every 1 hour. The uploaded data is encrypted to ensure data security. Cloud platform data processing and model fine-tuning: The cloud platform periodically summarizes and analyzes the uploaded historical data, focusing on extracting historical prediction deviation data. A joint optimization mechanism including model weight parameters and network structure hyperparameters is constructed. The specific fine-tuning and optimization process is as follows: On one hand, for the continuously differentiable weight matrix and bias vector within the LSTM model, gradient descent based on the Adam optimizer is used for fine-tuning. Using the minimization of the root mean square error (RMSE) of the prediction over the past 7 days as the objective function, gradients are calculated through backpropagation and network weights are updated, allowing the model to quickly adapt to recent data distribution shifts and improve short-term prediction accuracy. On the other hand, for non-differentiable discrete and probabilistic hyperparameters (including the number of hidden layer neurons, dropout coefficients, and initial learning rate), a Bayesian optimization algorithm is used for hyperparameter optimization. Since gradients cannot be computed in the hyperparameter space, the cloud platform constructs a Gaussian process surrogate model to evaluate the expected prediction error under different hyperparameter combinations and uses a sampling function (such as the expected improvement rate EI) to guide the sampling of the next set of hyperparameters. This mechanism can efficiently search for the optimal network topology and training configuration with limited cloud computing power, avoiding the computational waste caused by traditional grid search. The fine-tuned model parameters are sent to the local edge computing gateway via encrypted communication, overwriting the original model parameters for the next day's prediction work. Regarding the update of optimization algorithm strategy parameters, specifically: the cloud platform constructs a multi-dimensional weight adaptive allocation function and parameter deterministic mapping rules based on a two-dimensional seasonal-weather state, accurately sending the weight coefficients of the multi-objective optimization model and the strategy parameters of the improved particle swarm optimization algorithm. Let the weight of electricity cost be... The weight of local photovoltaic power consumption rate is And it satisfies the normalization constraint. The specific deterministic update rules are as follows: (1) Establish a two-dimensional state mapping table of season and weather, and determine the weight value according to the business-driven demand: Summer mode (high photovoltaic power generation, peak load, large peak-valley price difference, with economic efficiency as the main factor): The basic weight is strictly set as If the weather forecast predicts continuous cloudy and rainy days (leading to a sharp drop in photovoltaic output), a state switch will be triggered, and the system will be corrected. To ensure basic grid integration. Winter mode (low solar power generation, stable load, with grid integration as the primary focus): The basic load is strictly set at [value missing]. If the weather forecast indicates consecutive sunny days (and photovoltaic output exceeds expectations), a state switch will be triggered, and the system will be corrected to... To tap into economic potential. Spring / Autumn transition mode: adopting an equilibrium benchmark, set at... If extreme weather events (such as heavy rain or snow) cause the forecast confidence level to fall below a preset threshold, the system will be forcibly switched to an absolute balanced weighting. , to reduce scheduling risks. (2) The number of iterations is set by the deterministic mapping between the number of iterations of the particle swarm algorithm and the search strategy: the maximum number of iterations is set to 200 in summer mode to improve the optimization accuracy under complex working conditions; the maximum number of iterations is set to 100 in winter mode to match the computing power of edge devices and improve the solution speed; the number of iterations is set to 150 in spring / autumn mode. Inertial weight decay strategy: in sunny mode, the photovoltaic output is stable and the prediction error is small, so the inertial weight decay rate coefficient is set to 0.8 (slowing down the decay and maintaining local development capability); in cloudy and rainy mode, the photovoltaic output fluctuates greatly, so the inertial weight decay rate coefficient is set to 1.2. Feedback correction trigger sensitivity linkage: in cloudy and rainy mode, the relative error threshold trigger sensitivity in step S4 is increased by 20% to accelerate the response speed of macro boundary reconstruction. Through the above deterministic mapping rules, the cloud platform can accurately calculate and distribute weight parameters and algorithm configurations without relying on fuzzy logic, ensuring that the optimization target is strictly matched with the current actual working conditions. The formation of the closed-loop self-learning mechanism specifically includes: the cloud platform sends the updated algorithm strategy parameters (weight coefficients, iteration counts, inertia weights, etc.) to the local scheduling controller. The local controller combines the updated model and algorithm parameters to execute new prediction, optimization, and scheduling processes. At the same time, it continuously collects operational data and uploads it to the cloud platform, forming a closed-loop self-learning mechanism of local scheduling - data upload - cloud platform optimization - parameter distribution - local update. This ensures that the scheduling method can adapt to long-term changes in seasons, weather, and load, thereby improving the economy and stability of system operation.

[0052] like Figure 5 As shown, this embodiment also discloses an adaptive energy dispatch system for a distributed photovoltaic energy storage system, including: The data acquisition module is used to collect meteorological data of the distributed photovoltaic energy storage system based on meteorological sensors, combine it with the historical load curve and historical photovoltaic power generation obtained from the user-side smart meter to construct multi-source data, and preprocess the multi-source data. The prediction calculation module is used to input preprocessed multi-source data into an LSTM neural network model and output prediction results; and adopts a rolling time-domain update mechanism to periodically update the model input time series with the latest measured data and recalculate and output updated prediction results; wherein, the prediction results include a photovoltaic power generation prediction curve and a user load prediction curve for a preset period of time in the future; The optimization scheduling module is used to construct a multi-objective optimization model that includes minimizing electricity costs and maximizing local absorption rate, using the grid peak-valley flat electricity price, user load forecast, and energy storage battery state of charge as boundary conditions; the improved particle swarm optimization algorithm is used to solve the multi-objective optimization model and output the optimal charging and discharging power command sequence of the energy storage system within a preset time period in the future; The feedback control module is used by the dispatch controller to periodically compare the photovoltaic power generation forecast, user load forecast and corresponding measured values ​​to calculate the rolling forecast error. When the rolling forecast error exceeds the preset error threshold, feedback correction is triggered and the energy storage charging and discharging power limit and battery state of charge reference trajectory are dynamically adjusted for subsequent dispatch periods. The cloud platform management module is used to upload scheduling execution data, prediction error distribution, and optimization target achievement status to the cloud platform. The cloud platform regularly fine-tunes the parameters of the LSTM prediction model based on historical deviation data, and adaptively updates and improves the strategy parameters of the particle swarm algorithm according to different seasons and weather patterns, forming a closed-loop self-learning mechanism.

[0053] Specifically, the data acquisition module integrates a weather station, inverter communication interface, and smart meter, supporting mainstream industrial protocols; the prediction calculation module deploys an LSTM rolling prediction model, supporting lightweight inference at the edge and reducing dependence on gateway computing power; the optimization scheduling module incorporates a multi-objective optimization engine and an improved particle swarm solver, receiving prediction results and real-time electricity prices, and outputting scheduling instructions; the feedback control module includes an error monitoring unit and a SOC trajectory correction unit, achieving real-time policy smoothing and safety limiting; and the cloud platform management module is responsible for data aggregation, model version management, parameter iteration distribution, and visual review.

[0054] Taking a photovoltaic-storage system in an industrial and commercial park in Hainan as an example: the photovoltaic installed capacity is 1.2MW, and the energy storage configuration is 0.6MW / 1.2MWh.

[0055] Forecast window: 24 hours, time granularity: 15 minutes; LSTM model input includes features such as irradiance, temperature, humidity, wind speed, historical power, and load curve.

[0056] Optimize the solution: Improve the particle swarm optimization algorithm by reasonably configuring parameters such as population size and number of iterations based on the computing power of edge devices to ensure that a round of policy update is completed within 15 minutes.

[0057] Safety boundary: The default safety range of SOC is set to 20%~90%, which is automatically narrowed at low temperatures; the charging and discharging power command is sent to the PCS after being filtered.

[0058] Cloud platform iteration: Fine-tune model parameters weekly using recent execution data, and switch strategy parameter packages seasonally.

[0059] Actual trial operation showed that the monthly electricity cost in the park has decreased significantly, the local photovoltaic consumption rate has increased, and no battery safety alarms have occurred due to the dispatch strategy.

[0060] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An adaptive energy dispatch method for a distributed photovoltaic energy storage system, characterized in that, include: Meteorological data from distributed photovoltaic energy storage systems are collected using meteorological sensors. Combined with historical load curves and historical photovoltaic power generation obtained from user-side smart meters, multi-source data is constructed and preprocessed. The preprocessed multi-source data is input into the LSTM neural network model, and the prediction results are output. A rolling time-domain update mechanism is adopted, which periodically updates the model input time series with the latest measured data, recalculates and outputs the updated prediction results, and stores them in the local database; wherein, the prediction results include the photovoltaic power generation prediction curve and the user load prediction curve for a preset period of time in the future; Using the peak-valley flat electricity price of the power grid, user load forecasting, and the state of charge (SOC) of the energy storage battery as boundary conditions, a multi-objective optimization model is constructed, which includes minimizing electricity costs and maximizing local absorption rate. An improved particle swarm optimization algorithm is used to solve the multi-objective optimization model and output the optimal charging and discharging power command sequence of the energy storage system within a preset time period in the future. The dispatch controller periodically compares the predicted photovoltaic power generation, the predicted user load, and the corresponding measured values ​​to calculate the rolling prediction error. When the rolling prediction error exceeds the preset error threshold, feedback correction is triggered, and the energy storage charging and discharging power limits and battery state of charge reference trajectory are dynamically adjusted for subsequent dispatch periods. The scheduling execution data, prediction error distribution, and optimization target achievement status are uploaded to the cloud platform. The cloud platform regularly fine-tunes the parameters of the LSTM prediction model based on historical deviation data, and adaptively updates and improves the strategy parameters of the particle swarm algorithm according to different seasons and weather patterns, forming a closed-loop self-learning mechanism.

2. The adaptive energy dispatch method for a distributed photovoltaic energy storage system as described in claim 1, characterized in that, Preprocessing the multi-source data includes: Linear interpolation was used to align and correct the timestamps collected from different devices. The 3σ criterion combined with the Laida criterion is used to detect outliers in time series data, and to remove abnormal and invalid data caused by sudden changes in irradiance, load spikes, zero-value dead zones of photovoltaic power, and sensor fault drift. A scenario-based completion strategy is adopted to complete missing data. The preprocessed meteorological data, historical load sequences, and historical photovoltaic power sequences are reconstructed into a fixed-length model input time sample sequence in chronological order.

3. The adaptive energy dispatch method for a distributed photovoltaic energy storage system as described in claim 1, characterized in that, The LSTM prediction model includes an input layer, a hidden layer, and an output layer; The number of neurons in the input layer is consistent with the feature dimension of the preprocessed data. Two hidden layers are set, with 64 neurons in each layer, and the ReLU activation function is used; The output layer has two neurons, corresponding to the predicted photovoltaic power generation and the predicted user electricity load for a preset time period, respectively. The output layer uses a linear activation function.

4. The adaptive energy dispatch method for a distributed photovoltaic energy storage system as described in claim 1, characterized in that, A rolling time-domain update mechanism is adopted, which periodically updates the model input time series with the latest measured data and recalculates the updated prediction results, including: An update cycle consistent with the sampling period is adopted. The model input time series is updated periodically using the latest measured data. The oldest sampling point in the input series is removed, the latest measured sampling point is added, and the data is re-input into the LSTM prediction model for calculation, and the updated prediction result is output.

5. The adaptive energy dispatch method for a distributed photovoltaic energy storage system as described in claim 1, characterized in that, An improved particle swarm optimization algorithm is used to solve the multi-objective optimization model, outputting the optimal charging and discharging power command sequence for the energy storage system within a preset time period, including: S311: Initialize the particle swarm; S312: For the charge and discharge power sequence corresponding to each initial particle, perform two nonlinear constraints on battery charge and discharge efficiency and lifetime loss, calculate the actual charge and discharge power and calculate the loss value through the lifetime loss function, remove particles that do not meet the constraints, and reinitialize the particles. S313: Based on the comprehensive objective function, a lifetime loss penalty term is added to calculate the fitness value of each particle; S314: Based on the convergence judgment factor, determine the current state of the algorithm, adjust the inertia weight and learning factor, and update the particle velocity and position; S315: Determine if the iteration termination condition has been met: If it has, output the optimal position of the group, i.e. the optimal charge / discharge power command sequence; if it has not been met, return to step S312 to continue iterating until the termination condition is met.

6. The adaptive energy dispatch method for a distributed photovoltaic energy storage system as described in claim 1, characterized in that, The dispatch controller periodically compares the predicted photovoltaic power generation, the predicted user load, and the corresponding measured values ​​to calculate the rolling forecast error, including: The dispatch controller synchronously collects the measured values ​​of photovoltaic power and user load according to a preset cycle, and calls the corresponding photovoltaic power prediction values ​​and load prediction values ​​in the local database to calculate the rolling prediction error; the rolling prediction error calculation adopts a dual judgment of absolute error and relative error.

7. The adaptive energy dispatch method for a distributed photovoltaic energy storage system as described in claim 1, characterized in that, After the correction mechanism is triggered, the scheduling controller automatically adjusts the energy storage charging and discharging power limits and the battery state-of-charge reference trajectory for subsequent scheduling periods. Specific adjustment strategies include: If the predicted photovoltaic power value is consistently lower than the measured value, the limit for energy storage charging power will be increased while the limit for discharging power will be decreased to increase local photovoltaic consumption; if the predicted photovoltaic power value is consistently higher than the measured value, the limit for charging power will be decreased while the limit for discharging power will be increased. If the load forecast value is consistently higher than the measured value, the SOC reference trajectory for subsequent periods should be appropriately reduced to decrease the energy storage charging amount; if the load forecast value is consistently lower than the measured value, the SOC reference trajectory should be appropriately increased to increase the energy storage charging amount.

8. The adaptive energy dispatch method for a distributed photovoltaic energy storage system as described in claim 1, characterized in that, The cloud platform regularly fine-tunes the parameters of the LSTM prediction model based on historical bias data, including: The cloud platform regularly summarizes and analyzes the uploaded historical data, focusing on extracting historical prediction deviation data, and uses the gradient descent method to fine-tune the parameters of the LSTM prediction model.

9. The adaptive energy dispatch method for a distributed photovoltaic energy storage system as described in claim 1, characterized in that, Closed-loop self-learning mechanisms include: The cloud platform sends the updated algorithm strategy parameters to the local scheduling controller. The local controller combines the updated model and algorithm parameters to execute the new prediction, optimization and scheduling process. At the same time, it continuously collects running data and uploads it to the cloud platform, forming a closed-loop self-learning mechanism of local scheduling - data upload - cloud platform optimization - parameter distribution - local update.

10. An adaptive energy dispatching system for a distributed photovoltaic energy storage system, characterized in that, include: The data acquisition module is used to collect meteorological data of the distributed photovoltaic energy storage system based on meteorological sensors, combine it with the historical load curve and historical photovoltaic power generation obtained from the user-side smart meter to construct multi-source data, and preprocess the multi-source data. The prediction calculation module is used to input preprocessed multi-source data into the LSTM neural network model and output the prediction results. A rolling time-domain update mechanism is adopted, which periodically updates the model input time series with the latest measured data and recalculates and outputs the updated prediction results; wherein, the prediction results include the photovoltaic power generation prediction curve and the user load prediction curve for a preset period of time in the future; The optimization scheduling module is used to construct a multi-objective optimization model that includes minimizing electricity costs and maximizing local absorption rate, using the grid peak-valley flat electricity price, user load forecast, and energy storage battery state of charge as boundary conditions; the improved particle swarm optimization algorithm is used to solve the multi-objective optimization model and output the optimal charging and discharging power command sequence of the energy storage system within a preset time period in the future; The feedback control module is used by the dispatch controller to periodically compare the photovoltaic power generation forecast, user load forecast and corresponding measured values ​​to calculate the rolling forecast error. When the rolling forecast error exceeds the preset error threshold, feedback correction is triggered and the energy storage charging and discharging power limit and battery state of charge reference trajectory are dynamically adjusted for subsequent dispatch periods. The cloud platform management module is used to upload scheduling execution data, prediction error distribution, and optimization target achievement status to the cloud platform. The cloud platform regularly fine-tunes the parameters of the LSTM prediction model based on historical deviation data, and adaptively updates and improves the strategy parameters of the particle swarm algorithm according to different seasons and weather patterns, forming a closed-loop self-learning mechanism.

Images