A multi-source prediction-based energy storage scheduling method and system

Abstract

This invention provides a multi-source prediction-based energy storage dispatching method and system. The method includes: acquiring initial historical net load, measured photovoltaic data, real-time irradiance, initial real-time temperature, historical electricity price data, and real-time electricity price data; and acquiring predicted photovoltaic data, predicted net load, and predicted electricity price data based on the above data and a preset model to construct a multi-source prediction sequence; acquiring the real-time energy storage SOC trajectory and constructing a multi-objective function based on it and the above data; generating a simulated photovoltaic load curve based on the above data, and determining energy storage output data and SOC action threshold based on the multi-objective function, preset constraints, and the simulated photovoltaic load curve; and sending the above data to an energy storage converter to enable the energy storage converter to complete the energy storage dispatching of the power system. This invention provides a multi-source prediction-based energy storage dispatching method that avoids insufficient energy storage dispatching due to the influence of multi-source uncertainties, achieving dynamic optimization dispatching of energy storage in the power system.

Description

Technical Field

[0001] This invention relates to the field of new power system energy storage dispatch technology, and in particular to an energy storage dispatch method and system based on multi-source prediction. Background Technology

[0002] In the context of the current construction of new power systems characterized by a high proportion of renewable energy integration, user-side energy storage, as a key technology for improving grid flexibility, promoting clean energy consumption, and optimizing user energy costs, has become a research hotspot in both academia and industry for its refined management and value mining. Traditional energy storage control strategies often rely on preset fixed thresholds or time-of-use pricing ranges for simple "peak shaving and valley filling" operations. This rigid control strategy fails to fully consider the randomness of user load, the volatility of distributed photovoltaic output, and the spatiotemporal differences in electricity prices under the electricity market environment, resulting in the underutilization of the regulation potential and economic benefits of energy storage. With the development of artificial intelligence and big data technologies, prediction-based optimal scheduling has become possible, and existing technologies have begun to explore the introduction of predictions based on single factors (such as electricity prices or load). However, existing technologies still have significant limitations, lacking the collaborative perception and deep integration of multi-source uncertain information such as load, photovoltaics, and electricity prices, and often neglecting the health status of the battery energy storage itself in the optimization model, making it difficult to achieve an optimal balance between economic efficiency and equipment lifespan in the scheduling strategy.

[0003] Under the current technological background, energy storage dispatch schemes for power systems still have the following problems: First, existing strategies do not adequately consider the coordinated factors of multiple uncertainties. Most control models rely only on fixed time-of-use prices or single load thresholds, failing to deeply integrate multi-dimensional dynamic information such as short-term photovoltaic output, user load, and real-time electricity prices. This results in low matching degree between strategies and actual energy consumption scenarios, limiting optimization potential. Second, optimization models generally lack consideration of the long-term value of energy storage assets. Traditional optimization methods with intraday economic efficiency as the sole objective often ignore the degradation of battery cycle life caused by charging and discharging strategies. They lack a mechanism to quantify lifespan degradation as a cost and incorporate it into the objective function, which can easily lead to non-optimal returns over the entire life cycle. Finally, existing methods have poor adaptability and robustness in dealing with fluctuations. Their fixed-rule-based dispatch lacks proactiveness and cannot make rolling optimization decisions based on ultra-short-term forecasts. When faced with significant fluctuations in renewable energy output and market prices, the system response is lagging, the decision-making is rigid, and dynamic and fine-grained control cannot be achieved. Summary of the Invention

[0004] The present invention aims to provide a method and system for energy storage scheduling based on multi-source prediction, so as to solve the above-mentioned technical problems, avoid insufficient energy storage scheduling due to the influence of multi-source uncertainty factors, and realize dynamic optimization scheduling of power system energy storage.

[0005] To address the aforementioned technical problems, this invention provides a multi-source prediction-based energy storage scheduling method, comprising:

[0006] Acquire initial historical net load, measured photovoltaic data, real-time irradiance, initial real-time temperature, historical electricity price data, and real-time electricity price data;

[0007] Based on the initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model, and preset net load prediction model, the predicted photovoltaic data, predicted net load, and predicted electricity price data are obtained, and a multi-source prediction sequence is constructed based on the predicted photovoltaic data, predicted net load, and predicted electricity price data.

[0008] Obtain the real-time energy storage SOC trajectory, and construct a multi-objective function based on the real-time energy storage SOC trajectory, predicted photovoltaic data, and preset scheduling load;

[0009] Based on the initial historical net load, measured photovoltaic data, predicted photovoltaic data and predicted net load, several sets of simulated photovoltaic load curves are generated, and based on multi-objective functions, preset power balance constraints, preset battery state of charge constraints and simulated photovoltaic load curves, the energy storage output data and SOC action threshold are determined.

[0010] The multi-source prediction sequence, energy storage output data, and SOC action threshold are sent to the energy storage converter so that the energy storage converter can complete the energy storage scheduling of the power system based on the multi-source prediction sequence, energy storage output data, and SOC action threshold.

[0011] In the aforementioned scheme, by combining initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, and real-time electricity price data with preset photovoltaic output prediction models, preset net load prediction models, and preset electricity price prediction models, predicted photovoltaic data, predicted net load, and predicted electricity price data are generated, and a multi-source prediction sequence is constructed. This integrates multiple uncertainties related to photovoltaics, load, and electricity prices, providing high-precision dynamic information input for the optimized scheduling of energy storage in the power system. This avoids insufficient energy storage scheduling due to the influence of multiple uncertainties. Next, by acquiring the real-time energy storage SOC trajectory and combining it with predicted photovoltaic data and preset scheduling loads to construct a multi-objective function, the health status of energy storage devices and grid interaction needs can be incorporated into the optimization objectives. This breaks through the limitations of traditional single optimization and achieves a comprehensive consideration of the entire lifecycle value of energy storage scheduling. This provides a goal-oriented approach for subsequently acquiring energy storage output data and a scientific goal-oriented approach for accurately solving subsequent energy storage output data, further supporting the realization of optimized energy storage scheduling in the power system. Then, several sets of simulated photovoltaic load curves are generated using initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load. These curves are then combined with multi-objective functions, preset power balance constraints, and preset battery state of charge constraints to determine energy storage output data and SOC (State of Charge) action thresholds. This effectively handles the residual uncertainty arising from predictions based on multiple uncertainties, while transforming traditional fixed thresholds into dynamically adjustable SOC action thresholds that adapt to different scenarios. This enhances the robustness and adaptability of the power system's energy storage dispatch strategy, ensuring that energy storage dispatch for the power system can adapt to complex and ever-changing operating scenarios. Finally, by sending the multi-source prediction sequence, energy storage output data, and SOC action thresholds to the energy storage converter, the converter can accurately execute charge and discharge control based on these parameters, completing the energy storage dispatch for the power system and ensuring an effective transition from decision-making to execution.

[0012] Furthermore, the step of acquiring predicted photovoltaic data, predicted net load, and predicted electricity price data based on initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model, and preset net load prediction model, and constructing a multi-source prediction sequence based on the predicted photovoltaic data, predicted net load, and predicted electricity price data, includes:

[0013] Data cleaning is performed on the initial historical net load, real-time irradiance, and initial real-time temperature to obtain the first historical net load, moving average irradiance, and first real-time temperature.

[0014] Based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, predictive photovoltaic data is obtained.

[0015] Based on the predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the predicted net load is obtained.

[0016] Based on historical electricity price data, real-time electricity price data, and a preset electricity price prediction model, predictive electricity price data is obtained.

[0017] A multi-source prediction sequence is constructed based on predicted photovoltaic data, predicted net load, and predicted electricity price data.

[0018] In the above scheme, by cleaning the initial historical net load, real-time irradiance, and initial real-time temperature, outlier data can be removed and data features optimized to obtain the first historical net load, moving average irradiance, and first real-time temperature. This provides high-quality, high-reliability input data for subsequent prediction models, avoiding prediction bias caused by outlier data and ensuring prediction accuracy. Next, the predicted photovoltaic data is obtained by calculating using the moving average irradiance, the first real-time temperature, and a preset photovoltaic output prediction model, providing a basis for subsequent energy storage optimization and scheduling. Then, the predicted net load is obtained by using the predicted photovoltaic data, the first historical net load, and the preset net load prediction model. This explicitly models the nonlinear impact of photovoltaic output on the net load, solving the error superposition problem caused by traditional independent prediction and improving the accuracy of net load prediction. Subsequently, the predicted electricity price data is obtained by using historical electricity price data, real-time electricity price data, and a preset electricity price prediction model. This captures the spatiotemporal variation patterns and long-term dependencies of electricity prices, supporting the achievement of the goal of maximizing revenue throughout the entire life cycle. Finally, by integrating the predicted photovoltaic data, predicted net load, and predicted electricity price data, a multi-source prediction sequence is constructed. This enables the fusion of multi-source uncertainty information, providing comprehensive and dynamic information input for the dynamic optimization of subsequent energy storage scheduling, and avoiding mismatch between scheduling strategies and actual scenarios due to data fragmentation.

[0019] Furthermore, in the process of acquiring predicted photovoltaic data based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, the construction process of the preset photovoltaic output prediction model includes:

[0020] Obtain historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets;

[0021] Construct an initial photovoltaic power output prediction model;

[0022] Based on historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, the initial photovoltaic output prediction model is trained until the preset first convergence condition is met, thus obtaining the preset photovoltaic output prediction model.

[0023] In the above scheme, by acquiring historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, sufficient and realistic basic data can be provided for training the initial photovoltaic output prediction model, ensuring the effectiveness of the initial photovoltaic output prediction model training. Next, by constructing the initial photovoltaic output prediction model, a mapping framework between photovoltaic output, irradiance, and temperature is established, providing a basic architecture for subsequent optimization of the initial photovoltaic output prediction model and ensuring the feasibility of the prediction logic. Then, using the historical irradiance dataset, historical temperature dataset, and historical photovoltaic dataset, the initial photovoltaic output prediction model is iteratively trained until the preset first convergence condition is met. This allows the initial photovoltaic output prediction model to fully learn the inherent correlation between photovoltaic output and irradiance and temperature in historical data, ultimately obtaining a preset photovoltaic output prediction model with satisfactory accuracy and strong stability, providing reliable support for subsequent accurate photovoltaic data prediction.

[0024] Furthermore, in the process of obtaining the predicted net load based on predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the construction process of the preset net load prediction model includes:

[0025] Obtain historical photovoltaic datasets, historical net load datasets, and historical projected net load datasets;

[0026] Construct an initial net load forecasting model;

[0027] Based on historical photovoltaic datasets, historical net load datasets, and historical predicted net load datasets, the initial net load prediction model is trained until the preset second convergence condition is met, thus obtaining the preset net load prediction model.

[0028] In the above scheme, by acquiring historical photovoltaic (PV) datasets, historical net load datasets, and historical predicted net load datasets, comprehensive and realistic basic data can be provided for training the initial net load prediction model, ensuring the effectiveness and relevance of the initial net load prediction model training. Next, by constructing the initial net load prediction model, a mapping framework between PV output, net load, and predicted net load is established, providing an architecture for subsequent optimization of the initial net load prediction model. Then, using the historical PV dataset, historical net load dataset, and historical predicted net load dataset, the initial net load prediction model is iteratively trained until a preset second convergence condition is met. This allows the initial net load prediction model to fully learn the nonlinear coupling influence of PV output on net load in historical data and the temporal variation characteristics of net load itself. Ultimately, a preset net load prediction model with satisfactory accuracy and strong stability is obtained, providing reliable support for subsequent accurate output of predicted net load and avoiding the accumulation of errors caused by independent predictions.

[0029] Furthermore, the step of acquiring the real-time energy storage SOC trajectory and constructing a multi-objective function based on the real-time energy storage SOC trajectory, predicted photovoltaic data, and preset scheduling load includes:

[0030] The real-time energy storage SOC trajectory is obtained, and the discharge depth and cycle temperature of several charge-discharge cycles are extracted based on the real-time energy storage SOC trajectory.

[0031] A battery physical lifetime loss term is constructed based on depth of discharge and cycle temperature;

[0032] A grid interaction stability term is constructed based on predicted photovoltaic data and preset dispatch loads.

[0033] A multi-objective function is constructed based on the grid interaction stability term and the battery physical lifetime loss term.

[0034] In the above scheme, by acquiring the real-time SOC trajectory of the energy storage system and extracting the depth of discharge and cycle temperature of several charge-discharge cycles based on this trajectory, the core loss-influencing parameters of the energy storage device in actual operation can be accurately captured, providing data that fits the actual operating conditions for quantifying the physical life loss of the battery. Next, by constructing a battery physical life loss term using the depth of discharge and cycle temperature, the long-term cyclic degradation process of the battery can be transformed into a quantifiable cost indicator, thus providing a core quantitative basis for the optimization goal of maximizing the benefits throughout the entire life cycle. Then, by constructing a grid interaction stability term using predicted photovoltaic data and preset scheduling loads, the technical objective of the energy storage device in mitigating photovoltaic fluctuations and maintaining grid-side power stability can be quantified, avoiding grid interaction imbalance caused by photovoltaic output fluctuations. Subsequently, a multi-objective function is constructed using the grid interaction stability term and the battery physical life loss term, incorporating the grid operation stability requirements and the durability protection of the energy storage device into a unified optimization framework. This enables a synergistic consideration of economic efficiency, grid friendliness, and the long-term value of the equipment, providing a scientific and comprehensive objective guide for the formulation of dynamic optimization scheduling strategies for energy storage.

[0035] Furthermore, the process of generating several sets of simulated photovoltaic load curves based on initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load, and determining energy storage output data and SOC action thresholds based on multi-objective functions, preset power balance constraints, preset battery state of charge constraints, and simulated photovoltaic load curves, includes:

[0036] Based on the initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load, a probability density function for photovoltaic load prediction error is established, and several sets of simulated photovoltaic load curves are generated based on the probability density function for photovoltaic load prediction error.

[0037] With the goal of minimizing a multi-objective function, the simulated photovoltaic load curves are screened based on preset power balance constraints, preset battery state of charge constraints, and preset mathematical expectation conditions to obtain the first photovoltaic load curve. Then, based on the first photovoltaic load curve, the energy storage output data and the SOC action threshold are determined.

[0038] In the above scheme, a probability density function for photovoltaic load prediction error is established using initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load. Based on this probability density function, several sets of simulated photovoltaic load curves are randomly generated, which can comprehensively cover multi-source fluctuation scenarios of photovoltaics and loads. This transforms the uncertainty of prediction into a quantifiable set of scenarios, providing comprehensive scenario support for subsequent robustness optimization and avoiding insufficient fluctuation resistance of the scheduling strategy due to a single prediction scenario. Next, with the goal of minimizing a multi-objective function, the simulated photovoltaic load curves are screened by combining preset power balance constraints, preset battery state of charge constraints, and preset mathematical expectation conditions. This allows for the acquisition of a first photovoltaic load curve that balances grid stability and battery life, thereby determining the corresponding energy storage output data and the corresponding SOC action threshold. This ensures that the energy storage scheduling strategy, while meeting physical constraints, is both adaptable to multi-scenario fluctuations and maximizes the full life cycle value, improving the accuracy and adaptability of energy storage scheduling.

[0039] This invention provides an energy storage scheduling system based on multi-source prediction, comprising a multi-source data acquisition module, a multi-source sequence construction module, a multi-objective function construction module, a scheduling parameter solving module, and a scheduling execution module, specifically:

[0040] The multi-source data acquisition module is used to acquire initial historical net load, measured photovoltaic data, real-time irradiance, initial real-time temperature, historical electricity price data, and real-time electricity price data;

[0041] The multi-source sequence construction module is used to acquire predicted photovoltaic data, predicted net load and predicted electricity price data based on the initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model and preset net load prediction model, and to construct a multi-source prediction sequence based on the predicted photovoltaic data, predicted net load and predicted electricity price data;

[0042] The multi-objective function construction module is used to acquire the real-time energy storage SOC trajectory and construct a multi-objective function based on the real-time energy storage SOC trajectory, predicted photovoltaic data, and preset scheduling load.

[0043] The scheduling parameter solving module is used to generate several sets of simulated photovoltaic load curves based on the initial historical net load, measured photovoltaic data, predicted photovoltaic data and predicted net load, and to determine the energy storage output data and SOC action threshold based on the multi-objective function, preset power balance constraints, preset battery state of charge constraints and simulated photovoltaic load curves.

[0044] The scheduling execution module is used to send the multi-source prediction sequence, energy storage output data and SOC action threshold to the energy storage converter, so that the energy storage converter can complete the energy storage scheduling of the power system based on the multi-source prediction sequence, energy storage output data and SOC action threshold.

[0045] This invention provides an energy storage dispatching system based on multi-source prediction. In practical applications, it only requires a multi-source sequence construction module. By combining initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, and real-time electricity price data with preset photovoltaic output prediction models, preset net load prediction models, and preset electricity price prediction models, it generates predicted photovoltaic data, predicted net load, and predicted electricity price data and constructs a multi-source prediction sequence. This system can integrate multiple uncertainties in photovoltaic, load, and electricity price, providing high-precision dynamic information input for the optimized dispatching of energy storage in the power system. This avoids the influence of multiple uncertainties that could lead to insufficient energy storage dispatching in the power system. Next, a multi-objective function construction module is employed. By acquiring real-time energy storage SOC trajectories and combining them with predicted photovoltaic data and preset dispatch loads, a multi-objective function is constructed. This allows the health status of energy storage devices and grid interaction needs to be incorporated into the optimization objectives, breaking through the limitations of traditional single optimization. It achieves a comprehensive consideration of the entire lifecycle value of energy storage dispatch, providing a goal-oriented approach for subsequent acquisition of energy storage output data and a scientific goal-oriented approach for the accurate solution of subsequent energy storage output data, further supporting the realization of optimized energy storage dispatch in the power system. Then, a dispatch parameter solution module is used. Several sets of simulated photovoltaic load curves are generated using initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load. Combined with the multi-objective function, preset power balance constraints, and preset battery state of charge constraints, energy storage output data and SOC action thresholds are determined. This effectively handles the residual uncertainty caused by predictions based on multiple uncertain factors. At the same time, it transforms traditional fixed thresholds into SOC action thresholds that can be dynamically adjusted according to the scenario, improving the robustness and adaptability of the power system's energy storage dispatch strategy and ensuring that the energy storage dispatch of the power system can adapt to complex and ever-changing operating scenarios. Finally, a scheduling execution module is used to send multi-source prediction sequences, energy storage output data, and SOC action thresholds to the energy storage converter. This enables the energy storage converter to accurately execute charging and discharging control based on the multi-source prediction sequences, energy storage output data, and SOC action thresholds, thereby completing the energy storage scheduling of the power system and ensuring the effective transformation from decision-making to execution.

[0046] Further, the multi-source sequence construction module is used to acquire predicted photovoltaic data, predicted net load, and predicted electricity price data based on initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model, and preset net load prediction model, and to construct a multi-source prediction sequence based on the predicted photovoltaic data, predicted net load, and predicted electricity price data, including:

[0047] Data cleaning is performed on the initial historical net load, real-time irradiance, and initial real-time temperature to obtain the first historical net load, moving average irradiance, and first real-time temperature.

[0048] Based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, predictive photovoltaic data is obtained.

[0049] Based on the predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the predicted net load is obtained.

[0050] Based on historical electricity price data, real-time electricity price data, and a preset electricity price prediction model, predictive electricity price data is obtained.

[0051] A multi-source prediction sequence is constructed based on predicted photovoltaic data, predicted net load, and predicted electricity price data.

[0052] In the above scheme, by cleaning the initial historical net load, real-time irradiance, and initial real-time temperature, outlier data can be removed and data features optimized to obtain the first historical net load, moving average irradiance, and first real-time temperature. This provides high-quality, high-reliability input data for subsequent prediction models, avoiding prediction bias caused by outlier data and ensuring prediction accuracy. Next, the predicted photovoltaic data is obtained by calculating using the moving average irradiance, the first real-time temperature, and a preset photovoltaic output prediction model, providing a basis for subsequent energy storage optimization and scheduling. Then, the predicted net load is obtained by using the predicted photovoltaic data, the first historical net load, and the preset net load prediction model. This explicitly models the nonlinear impact of photovoltaic output on the net load, solving the error superposition problem caused by traditional independent prediction and improving the accuracy of net load prediction. Subsequently, the predicted electricity price data is obtained by using historical electricity price data, real-time electricity price data, and a preset electricity price prediction model. This captures the spatiotemporal variation patterns and long-term dependencies of electricity prices, supporting the achievement of the goal of maximizing revenue throughout the entire life cycle. Finally, by integrating the predicted photovoltaic data, predicted net load, and predicted electricity price data, a multi-source prediction sequence is constructed. This enables the fusion of multi-source uncertainty information, providing comprehensive and dynamic information input for the dynamic optimization of subsequent energy storage scheduling, and avoiding mismatch between scheduling strategies and actual scenarios due to data fragmentation.

[0053] Furthermore, in the process of acquiring predicted photovoltaic data based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, the construction process of the preset photovoltaic output prediction model includes:

[0054] Obtain historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets;

[0055] Construct an initial photovoltaic power output prediction model;

[0056] Based on historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, the initial photovoltaic output prediction model is trained until the preset first convergence condition is met, thus obtaining the preset photovoltaic output prediction model.

[0057] In the above scheme, by acquiring historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, sufficient and realistic basic data can be provided for training the initial photovoltaic output prediction model, ensuring the effectiveness of the initial photovoltaic output prediction model training. Next, by constructing the initial photovoltaic output prediction model, a mapping framework between photovoltaic output, irradiance, and temperature is established, providing a basic architecture for subsequent optimization of the initial photovoltaic output prediction model and ensuring the feasibility of the prediction logic. Then, using the historical irradiance dataset, historical temperature dataset, and historical photovoltaic dataset, the initial photovoltaic output prediction model is iteratively trained until the preset first convergence condition is met. This allows the initial photovoltaic output prediction model to fully learn the inherent correlation between photovoltaic output and irradiance and temperature in historical data, ultimately obtaining a preset photovoltaic output prediction model with satisfactory accuracy and strong stability, providing reliable support for subsequent accurate photovoltaic data prediction.

[0058] Furthermore, in the process of obtaining the predicted net load based on predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the construction process of the preset net load prediction model includes:

[0059] Obtain historical photovoltaic datasets, historical net load datasets, and historical projected net load datasets;

[0060] Construct an initial net load forecasting model;

[0061] Based on historical photovoltaic datasets, historical net load datasets, and historical predicted net load datasets, the initial net load prediction model is trained until the preset second convergence condition is met, thus obtaining the preset net load prediction model.

[0062] In the above scheme, by acquiring historical photovoltaic (PV) datasets, historical net load datasets, and historical predicted net load datasets, comprehensive and realistic basic data can be provided for training the initial net load prediction model, ensuring the effectiveness and relevance of the initial net load prediction model training. Next, by constructing the initial net load prediction model, a mapping framework between PV output, net load, and predicted net load is established, providing an architecture for subsequent optimization of the initial net load prediction model. Then, using the historical PV dataset, historical net load dataset, and historical predicted net load dataset, the initial net load prediction model is iteratively trained until a preset second convergence condition is met. This allows the initial net load prediction model to fully learn the nonlinear coupling influence of PV output on net load in historical data and the temporal variation characteristics of net load itself. Ultimately, a preset net load prediction model with satisfactory accuracy and strong stability is obtained, providing reliable support for subsequent accurate output of predicted net load and avoiding the accumulation of errors caused by independent predictions. Attached Figure Description

[0063] Figure 1 A flowchart illustrating a multi-source prediction-based energy storage scheduling method according to an embodiment of the present invention;

[0064] Figure 2 This is an architecture diagram of an energy storage scheduling system based on multi-source prediction, provided as an embodiment of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] This embodiment provides an energy storage scheduling method based on multi-source prediction. Please refer to the flowchart for the method. Figure 1 ,include:

[0067] Step S1: Obtain initial historical net load, measured photovoltaic data, real-time irradiance, initial real-time temperature, historical electricity price data, and real-time electricity price data;

[0068] Step S2: Based on the initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model, and preset net load prediction model, obtain the predicted photovoltaic data, predicted net load, and predicted electricity price data, and construct a multi-source prediction sequence based on the predicted photovoltaic data, predicted net load, and predicted electricity price data;

[0069] Step S3: Obtain the real-time energy storage SOC trajectory, and construct a multi-objective function based on the real-time energy storage SOC trajectory, predicted photovoltaic data, and preset scheduling load;

[0070] Step S4: Based on the initial historical net load, measured photovoltaic data, predicted photovoltaic data and predicted net load, generate several sets of simulated photovoltaic load curves, and determine the energy storage output data and SOC action threshold based on the multi-objective function, preset power balance constraints, preset battery state of charge constraints and simulated photovoltaic load curves.

[0071] Step S5: Send the multi-source prediction sequence, energy storage output data and SOC action threshold to the energy storage converter so that the energy storage converter can complete the energy storage scheduling of the power system based on the multi-source prediction sequence, energy storage output data and SOC action threshold.

[0072] This embodiment focuses on a typical industrial and commercial user-side energy storage application scenario within a large science and technology park in a coastal city. The park has a high photovoltaic penetration rate, significant load fluctuations, and implements a time-of-use pricing policy. Specifically, this embodiment targets a research and development center with an annual electricity consumption of approximately 3.6 million kWh and an installed capacity of 300 kWp distributed photovoltaic system, equipped with a 500 kW / 1000 kWh lithium iron phosphate battery energy storage system. In this embodiment, please refer to Table 1. The initial historical net load is obtained through the smart meters (accuracy level 0.5S) in the power distribution room of the park. The real-time irradiance and initial real-time temperature are obtained through the weather station on the photovoltaic array side. The historical electricity price data and real-time electricity price data are obtained through the data interface interconnected with the power grid dispatch center. The measured photovoltaic data is obtained through the photovoltaic inverter. At the same time, the time information corresponding to the data acquisition is obtained according to the timestamp unit built into each acquisition device. The initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data and real-time electricity price data are combined with the preset photovoltaic output prediction model, preset net load prediction model and preset electricity price prediction model to generate predicted photovoltaic data, predicted net load and predicted electricity price data and construct a multi-source prediction sequence. This can integrate the uncertainties of photovoltaic, load and electricity price, and provide high-precision dynamic information input for the energy storage optimization and scheduling of the power system, thereby avoiding the influence of multiple uncertainties and the resulting insufficient energy storage scheduling of the power system.

[0073] Table 1 Multi-source data acquisition specifications

[0074]

[0075] Next, by acquiring real-time energy storage SOC trajectories and constructing a multi-objective function in conjunction with predicted photovoltaic data and preset dispatch loads, the health status of energy storage devices and grid interaction needs can be incorporated into the optimization objectives. This breaks through the limitations of traditional single optimization and achieves a comprehensive consideration of the entire lifecycle value of energy storage dispatch. This provides a goal-oriented approach for subsequently acquiring energy storage output data and a scientific goal-oriented approach for accurately solving subsequent energy storage output data, further supporting the realization of optimized energy storage dispatch in the power system. Then, several sets of simulated photovoltaic load curves are generated using initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load. Combined with the multi-objective function, preset power balance constraints, and preset battery state of charge constraints, energy storage output data and SOC action thresholds are determined. This effectively handles the residual uncertainty caused by predictions based on multiple sources of uncertainty, and transforms traditional fixed thresholds into SOC action thresholds that can be dynamically adjusted according to the scenario. This improves the robustness and adaptability of the power system's energy storage dispatch strategy, ensuring that energy storage dispatch in the power system can adapt to complex and ever-changing operating scenarios. The preset power balance constraint is... ,in, This represents the real-time power exchange between the user side and the power grid. Data on energy storage output, To predict net load, when When the user purchases electricity from the grid, the grid supplies electricity to the user; when At this time, when users sell electricity to the grid, a situation occurs where electricity is fed back into the grid; when At that time, the energy storage device was not in operation. Finally, the multi-source prediction sequence, energy storage output data, and SOC action threshold were sent to the energy storage converter (PCS) via the Modbus-TCP protocol. The closed-loop PID controller in the energy storage converter accurately tracked and rapidly adjusted the charging and discharging power, ensuring the effective execution of the dynamic threshold. This enabled the energy storage converter to accurately execute charging and discharging control based on the multi-source prediction sequence, energy storage output data, and SOC action threshold, completing the energy storage dispatch of the power system and ensuring the effective transformation from decision-making to execution. Furthermore, to achieve optimal control under complex operating conditions, the energy storage converter also has a built-in scene recognition unit based on fuzzy logic. This unit can analyze the grid dispatch instructions corresponding to the multi-source prediction sequence and energy storage output data in real time, automatically identifying the current typical operating mode. Its mode characteristics and control strategies are shown in Table 2 below. For example, when the "PV self-consumption" mode is identified, the energy storage device will prioritize maintaining the SOC in a lower range to maximize the absorption of PV power; while when switching to the "demand response" mode, it will quickly increase the upper limit of the discharge power to respond to grid dispatch. This multi-scenario adaptive mechanism significantly enhances the smooth switching capability and global adaptability of energy storage devices between different operating objectives through online fine-tuning of control parameters, ensuring the effectiveness and safety of dynamic strategies in real engineering environments.

[0076] Table 2 Correspondence between Multi-Scenario Operation Modes and Control Strategies

[0077]

[0078] Furthermore, the step of acquiring predicted photovoltaic data, predicted net load, and predicted electricity price data based on initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model, and preset net load prediction model, and constructing a multi-source prediction sequence based on the predicted photovoltaic data, predicted net load, and predicted electricity price data, includes:

[0079] Data cleaning is performed on the initial historical net load, real-time irradiance, and initial real-time temperature to obtain the first historical net load, moving average irradiance, and first real-time temperature.

[0080] Based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, predictive photovoltaic data is obtained.

[0081] Based on the predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the predicted net load is obtained.

[0082] Based on historical electricity price data, real-time electricity price data, and a preset electricity price prediction model, predictive electricity price data is obtained.

[0083] A multi-source prediction sequence is constructed based on predicted photovoltaic data, predicted net load, and predicted electricity price data.

[0084] In this embodiment, by cleaning the initial historical net load, real-time irradiance, and initial real-time temperature, outlier data can be removed and data features optimized to obtain the first historical net load, moving average irradiance, and first real-time temperature. This provides high-quality, highly reliable input data for subsequent prediction models, avoiding prediction bias caused by outlier data and ensuring prediction accuracy. Next, the predicted photovoltaic data is obtained by calculating the moving average irradiance I(t), the first real-time temperature T(t), and the preset photovoltaic output prediction model. This provides a predictive basis for subsequent energy storage optimization and scheduling. Then, the obtained predictive photovoltaic data... As a known prior feature vector, it is used to construct the first historical net load sequence with the first historical net load. Simultaneously, the time information corresponding to the data collected as described above. The data is then concatenated and input into a preset net load prediction model to obtain the predicted net load. Specifically: ,in, To predict net load, the nonlinear impact of photovoltaic (PV) output on net load was explicitly modeled through the above process, resolving the error superposition problem caused by traditional independent prediction. This allows the LSTM network to learn the nonlinear shading or peak-shaving effect of PV output changes on the net load curve during training through a gating mechanism, improving the accuracy of net load prediction. Subsequently, predicted electricity price data was obtained through historical electricity price data, real-time electricity price data, and a preset electricity price prediction model. The preset electricity price prediction model adopts a Transformer model, which can capture the spatiotemporal variation patterns and long-term dependencies of electricity prices, supporting the subsequent achievement of the goal of maximizing revenue over the entire life cycle. All the above models are executed in a rolling 15-minute cycle to generate predicted PV data, predicted net load, and predicted electricity price data for the next 4 hours (N=16 points). Finally, by integrating the predicted PV data, predicted net load, and predicted electricity price data, a multi-source prediction sequence is constructed. This allows for the fusion of multi-source uncertainty information, providing comprehensive and dynamic information input for the dynamic optimization of subsequent energy storage scheduling, avoiding mismatch between scheduling strategies and actual scenarios due to data fragmentation.

[0085] Furthermore, in the process of acquiring predicted photovoltaic data based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, the construction process of the preset photovoltaic output prediction model includes:

[0086] Obtain historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets;

[0087] Construct an initial photovoltaic power output prediction model;

[0088] Based on historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, the initial photovoltaic output prediction model is trained until the preset first convergence condition is met, thus obtaining the preset photovoltaic output prediction model.

[0089] In this embodiment, by acquiring historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, sufficient and realistic basic data can be provided for training the initial photovoltaic output prediction model, ensuring the effectiveness of the initial photovoltaic output prediction model training. Next, by constructing the initial photovoltaic output prediction model, a mapping framework between photovoltaic output, irradiance, and temperature is established, providing a basic architecture for subsequent optimization of the initial photovoltaic output prediction model and ensuring the feasibility of the prediction logic. Then, using the historical irradiance dataset, historical temperature dataset, and historical photovoltaic dataset, the initial photovoltaic output prediction model is iteratively trained until a preset first convergence condition is met. This allows the initial photovoltaic output prediction model to fully learn the inherent correlation between photovoltaic output and irradiance and temperature in historical data, ultimately obtaining a preset photovoltaic output prediction model with satisfactory accuracy and strong stability, providing reliable support for subsequent accurate photovoltaic data prediction.

[0090] Furthermore, in the process of obtaining the predicted net load based on predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the construction process of the preset net load prediction model includes:

[0091] Obtain historical photovoltaic datasets, historical net load datasets, and historical projected net load datasets;

[0092] Construct an initial net load forecasting model;

[0093] Based on historical photovoltaic datasets, historical net load datasets, and historical predicted net load datasets, the initial net load prediction model is trained until the preset second convergence condition is met, thus obtaining the preset net load prediction model.

[0094] In this embodiment, by acquiring historical photovoltaic (PV) datasets, historical net load datasets, and historical predicted net load datasets, comprehensive and realistic basic data can be provided for training the initial net load prediction model, ensuring the effectiveness and relevance of the initial net load prediction model training. Next, an initial net load prediction model is constructed using a Long Short-Term Memory (LSTM) network, establishing a mapping framework between PV output, net load, and predicted net load, providing an architecture for subsequent optimization of the initial net load prediction model. Then, using the historical PV dataset, historical net load dataset, and historical predicted net load dataset, the initial net load prediction model is iteratively trained until a preset second convergence condition is met. This allows the initial net load prediction model to fully learn the nonlinear coupling influence of PV output on net load in historical data and the temporal variation characteristics of net load itself. Ultimately, a preset net load prediction model with satisfactory accuracy and strong stability is obtained, providing reliable support for subsequent accurate output of predicted net load and avoiding the accumulation of errors caused by independent predictions.

[0095] Furthermore, the step of acquiring the real-time energy storage SOC trajectory and constructing a multi-objective function based on the real-time energy storage SOC trajectory, predicted photovoltaic data, and preset scheduling load includes:

[0096] The real-time energy storage SOC trajectory is obtained, and the discharge depth and cycle temperature of several charge-discharge cycles are extracted based on the real-time energy storage SOC trajectory.

[0097] A battery physical lifetime loss term is constructed based on depth of discharge and cycle temperature;

[0098] A grid interaction stability term is constructed based on predicted photovoltaic data and preset dispatch loads.

[0099] A multi-objective function is constructed based on the grid interaction stability term and the battery physical lifetime loss term.

[0100] In this embodiment, Rainflow (rainflow counting method) and Miner's linear cumulative damage theory are introduced. By acquiring the real-time energy storage SOC trajectory, the complex continuous charge and discharge process is transformed into a series of discrete cyclic events based on this real-time energy storage SOC trajectory, and the discharge depth of several charge and discharge cycles is extracted. and cycling temperature This allows for the precise capture of key loss-influencing parameters in the actual operation of energy storage devices, providing data closely aligned with real-world operating conditions for quantifying battery physical life loss. Next, a battery physical life loss term is constructed using depth of discharge and cycle temperature, specifically... Where M is the total number of cycles identified within the prediction period; Let be the depth of discharge in the k-th cycle; This represents the marginal physical loss rate caused by a single cycle. The Ncycle function is usually fitted using a power function form (e.g., ...). Where N represents the theoretical maximum cycle life under specific operating conditions, DoD represents the depth of discharge in a single charge-discharge cycle, and A and B are battery characteristic fitting constants determined by the specific battery type and experimental data. This process transforms the long-term cycle degradation process of the battery into a quantifiable cost indicator, thus providing a core quantitative basis for maximizing the optimization objective of the entire lifecycle. Then, by predicting photovoltaic data and pre-set dispatch loads to construct a grid interaction stability term, the technical objective of energy storage devices in mitigating photovoltaic fluctuations and maintaining grid-side power stability can be quantified, avoiding grid interaction imbalances caused by photovoltaic output fluctuations. Subsequently, a multi-objective function is constructed using the grid interaction stability term and the battery physical lifespan loss term, specifically:

[0101] ;

[0102] in, For the grid interaction stability term, by minimizing the sum of squared Euclidean distances between the grid exchange power and the benchmark value, the technical capability of energy storage devices to mitigate random fluctuations in photovoltaic power and maintain grid stability is quantified. , The weighting coefficients for each phase can be adaptively adjusted based on actual battery degradation data and changes in market rules. To preset the dispatch load, it is usually set to the average net load or dispatch command value within the forecast period. As a battery physical lifespan loss item, it incorporates the grid operation stability requirements and energy storage equipment durability protection into a unified optimization framework, enabling a synergistic consideration of economic efficiency, grid friendliness, and long-term equipment value, and providing a scientific and comprehensive goal orientation for the formulation of dynamic optimization scheduling strategies for energy storage.

[0103] Furthermore, the process of generating several sets of simulated photovoltaic load curves based on initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load, and determining energy storage output data and SOC action thresholds based on multi-objective functions, preset power balance constraints, preset battery state of charge constraints, and simulated photovoltaic load curves, includes:

[0104] Based on the initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load, a probability density function for photovoltaic load prediction error is established, and several sets of simulated photovoltaic load curves are generated based on the probability density function for photovoltaic load prediction error.

[0105] With the goal of minimizing a multi-objective function, the simulated photovoltaic load curves are screened based on preset power balance constraints, preset battery state of charge constraints, and preset mathematical expectation conditions to obtain the first photovoltaic load curve. Then, based on the first photovoltaic load curve, the energy storage output data and the SOC action threshold are determined.

[0106] In this embodiment, a probability density function for photovoltaic load prediction error is established using initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load. Based on this probability density function, Monte Carlo simulation is used to randomly generate several sets (e.g., 1000 sets) of simulated photovoltaic load curves. ,in For the simulated photovoltaic load curve of group s, where S can be 1000 groups, the process is as follows: The probability density function of the photovoltaic load prediction error generates 1000 groups of photovoltaic output errors and 1000 groups of load errors. These 1000 groups of photovoltaic output errors are then superimposed on the predicted photovoltaic data to obtain 1000 distinct simulated photovoltaic output curves. Similarly, these 1000 groups of load errors are superimposed on the predicted net load to obtain 1000 distinct simulated load curves. Finally, the 1000 distinct simulated photovoltaic output curves and 1000 distinct simulated load curves are combined to generate the simulated photovoltaic load curve. This allows for comprehensive coverage of multi-source fluctuation scenarios involving photovoltaics and loads, transforming the uncertainty of predictions into a quantifiable set of scenarios. This provides comprehensive scenario support for subsequent robust optimization, avoiding insufficient fluctuation resistance of the scheduling strategy due to a single prediction scenario. Next, with the goal of minimizing a multi-objective function, the simulated photovoltaic load curves are screened by combining preset power balance constraints, preset battery state of charge constraints, and preset mathematical expectation conditions. This yields a first photovoltaic load curve that balances grid stability and battery lifespan. This first photovoltaic load curve has the smallest mathematical expectation value among the aforementioned sets of simulated photovoltaic load curves. Based on this first photovoltaic load curve, the energy storage output data and corresponding SOC action threshold for the next cycle are determined, forming a series of conditional triggering rules that adaptively adjust according to the prediction scenario, as shown in Table 3. This ensures that the energy storage scheduling strategy, while meeting physical constraints, adapts to multi-scenario fluctuations and maximizes its full lifecycle value, improving the accuracy and adaptability of energy storage scheduling. The mathematical expectation value can be calculated using the following formula: The preset battery state of charge constraint is SOC. min <SOC(t)<SOC max Here, SOC(t) represents the remaining battery capacity of the energy storage device, which is related to the energy storage output data. Changes in the energy storage output data will directly lead to an increase or decrease in SOC(t). min The minimum state of charge (SOC) threshold for batteries in energy storage devices. max This is the highest state of charge threshold for the battery of the energy storage device.

[0107] Table 3 Examples of Key Parameters for Dynamic Threshold Strategy

[0108]

[0109] This embodiment provides an energy storage dispatch system based on multi-source prediction. Please refer to [link to relevant documentation]. Figure 2 It includes a multi-source data acquisition module, a multi-source sequence construction module, a multi-objective function construction module, a scheduling parameter solving module, and a scheduling execution module, specifically:

[0110] The multi-source data acquisition module is used to acquire initial historical net load, measured photovoltaic data, real-time irradiance, initial real-time temperature, historical electricity price data, and real-time electricity price data;

[0111] The multi-source sequence construction module is used to acquire predicted photovoltaic data, predicted net load and predicted electricity price data based on the initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model and preset net load prediction model, and to construct a multi-source prediction sequence based on the predicted photovoltaic data, predicted net load and predicted electricity price data;

[0112] The multi-objective function construction module is used to acquire the real-time energy storage SOC trajectory and construct a multi-objective function based on the real-time energy storage SOC trajectory, predicted photovoltaic data, and preset scheduling load.

[0113] The scheduling parameter solving module is used to generate several sets of simulated photovoltaic load curves based on the initial historical net load, measured photovoltaic data, predicted photovoltaic data and predicted net load, and to determine the energy storage output data and SOC action threshold based on the multi-objective function, preset power balance constraints, preset battery state of charge constraints and simulated photovoltaic load curves.

[0114] The scheduling execution module is used to send the multi-source prediction sequence, energy storage output data and SOC action threshold to the energy storage converter, so that the energy storage converter can complete the energy storage scheduling of the power system based on the multi-source prediction sequence, energy storage output data and SOC action threshold.

[0115] This embodiment provides an energy storage dispatch system based on multi-source prediction, targeting a typical industrial and commercial user-side energy storage application scenario within a large science and technology park in a coastal city. This park has a high photovoltaic penetration rate, significant load fluctuations, and implements a time-of-use pricing policy. Specifically, this embodiment applies to a research and development center with an annual electricity consumption of approximately 3.6 million kWh and an installed capacity of 300 kWp distributed photovoltaic power, equipped with a 500 kW / 1000 kWh lithium iron phosphate battery energy storage system. In this embodiment, only a multi-source data acquisition module is needed in practical applications. The initial historical net load is obtained through smart meters (accuracy level 0.5S) in the power distribution room of the park, real-time irradiance and initial real-time temperature are obtained through the weather station on the photovoltaic array side, historical electricity price data and real-time electricity price data are obtained through the data interface interconnected with the power grid dispatch center, and measured photovoltaic data is obtained through the photovoltaic inverter. At the same time, the time information corresponding to the data acquisition is obtained according to the timestamp unit built into each acquisition device. The multi-source sequence construction module combines the initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data and real-time electricity price data with the preset photovoltaic output prediction model, preset net load prediction model and preset electricity price prediction model to generate predicted photovoltaic data, predicted net load and predicted electricity price data and construct a multi-source prediction sequence. This can integrate the uncertainties of photovoltaic, load and electricity price from multiple sources, and provide high-precision dynamic information input for the energy storage optimization and scheduling of the power system, thereby avoiding the influence of multiple uncertainties that lead to insufficient energy storage scheduling of the power system. Next, a multi-objective function construction module is employed. By acquiring real-time energy storage SOC trajectories and combining them with predicted photovoltaic data and preset dispatch loads, a multi-objective function is constructed. This incorporates the health status of energy storage devices and grid interaction requirements into the optimization objectives, breaking through the limitations of traditional single optimization. It achieves a comprehensive consideration of the entire lifecycle value of energy storage dispatch, providing a goal-oriented approach for subsequent acquisition of energy storage output data and a scientific goal-oriented approach for the accurate solution of subsequent energy storage output data, further supporting the realization of optimized energy storage dispatch in the power system. Then, a dispatch parameter solution module is used. Several sets of simulated photovoltaic load curves are generated using initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load. Combined with the multi-objective function, preset power balance constraints, and preset battery state of charge constraints, energy storage output data and SOC action thresholds are determined. This effectively handles the residual uncertainty caused by predictions based on multiple sources of uncertainty, while transforming traditional fixed thresholds into SOC action thresholds that can be dynamically adjusted according to the scenario. This improves the robustness and adaptability of the power system's energy storage dispatch strategy, ensuring that energy storage dispatch in the power system can adapt to complex and ever-changing operating scenarios. The preset power balance constraint is... ,in, This represents the real-time power exchange between the user side and the power grid. Data on energy storage output, To predict net load, when When the user purchases electricity from the grid, the grid supplies electricity to the user; when At this time, when users sell electricity to the grid, a situation occurs where electricity is fed back into the grid; when At that time, the energy storage device was not in operation. Finally, a scheduling execution module was used to send the multi-source prediction sequence, energy storage output data, and SOC action threshold to the energy storage converter (PCS) via the Modbus-TCP protocol. The closed-loop PID controller in the energy storage converter accurately tracks and rapidly adjusts the charging and discharging power, ensuring the effective execution of the dynamic threshold. This allows the energy storage converter to precisely execute charging and discharging control based on the multi-source prediction sequence, energy storage output data, and SOC action threshold, completing the energy storage scheduling of the power system and ensuring an effective transition from decision-making to execution. Furthermore, to achieve optimal control under complex operating conditions, the energy storage converter also incorporates a fuzzy logic-based scene recognition unit. This unit can analyze the grid dispatch instructions corresponding to the multi-source prediction sequence and energy storage output data in real time, automatically determining the current typical operating mode. For example, when it identifies the "PV self-consumption" mode, the energy storage device will prioritize maintaining the SOC in a lower range to maximize PV absorption; while when switching to the "demand response" mode, it will rapidly increase the discharge power limit to respond to grid dispatch. This multi-scenario adaptive mechanism significantly enhances the smooth switching capability and global adaptability of energy storage devices between different operating objectives through online fine-tuning of control parameters, ensuring the effectiveness and safety of dynamic strategies in real engineering environments.

[0116] Further, the multi-source sequence construction module is used to acquire predicted photovoltaic data, predicted net load, and predicted electricity price data based on initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model, and preset net load prediction model, and to construct a multi-source prediction sequence based on the predicted photovoltaic data, predicted net load, and predicted electricity price data, including:

[0117] Data cleaning is performed on the initial historical net load, real-time irradiance, and initial real-time temperature to obtain the first historical net load, moving average irradiance, and first real-time temperature.

[0118] Based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, predictive photovoltaic data is obtained.

[0119] Based on the predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the predicted net load is obtained.

[0120] Based on historical electricity price data, real-time electricity price data, and a preset electricity price prediction model, predictive electricity price data is obtained.

[0121] A multi-source prediction sequence is constructed based on predicted photovoltaic data, predicted net load, and predicted electricity price data.

[0122] In this embodiment, by cleaning the initial historical net load, real-time irradiance, and initial real-time temperature, outlier data can be removed and data features optimized to obtain the first historical net load, moving average irradiance, and first real-time temperature. This provides high-quality, highly reliable input data for subsequent prediction models, avoiding prediction bias caused by outlier data and ensuring prediction accuracy. Next, the predicted photovoltaic data is obtained by calculating the moving average irradiance I(t), the first real-time temperature T(t), and the preset photovoltaic output prediction model. This provides a predictive basis for subsequent energy storage optimization and scheduling. Then, the obtained predictive photovoltaic data... As a known prior feature vector, it is used to construct the first historical net load sequence with the first historical net load. Simultaneously, the time information corresponding to the data collected as described above. The data is then concatenated and input into a preset net load prediction model to obtain the predicted net load. Specifically: ,in, To predict net load, the nonlinear impact of photovoltaic (PV) output on net load was explicitly modeled through the above process, resolving the error superposition problem caused by traditional independent prediction. This allows the LSTM network to learn the nonlinear shading or peak-shaving effect of PV output changes on the net load curve during training through a gating mechanism, improving the accuracy of net load prediction. Subsequently, predicted electricity price data was obtained through historical electricity price data, real-time electricity price data, and a preset electricity price prediction model. The preset electricity price prediction model adopts a Transformer model, which can capture the spatiotemporal variation patterns and long-term dependencies of electricity prices, supporting the subsequent achievement of the goal of maximizing revenue over the entire life cycle. All the above models are executed in a rolling 15-minute cycle to generate predicted PV data, predicted net load, and predicted electricity price data for the next 4 hours (N=16 points). Finally, by integrating the predicted PV data, predicted net load, and predicted electricity price data, a multi-source prediction sequence is constructed. This allows for the fusion of multi-source uncertainty information, providing comprehensive and dynamic information input for the dynamic optimization of subsequent energy storage scheduling, avoiding mismatch between scheduling strategies and actual scenarios due to data fragmentation.

[0123] Furthermore, in the process of acquiring predicted photovoltaic data based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, the construction process of the preset photovoltaic output prediction model includes:

[0124] Obtain historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets;

[0125] Construct an initial photovoltaic power output prediction model;

[0126] Based on historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, the initial photovoltaic output prediction model is trained until the preset first convergence condition is met, thus obtaining the preset photovoltaic output prediction model.

[0127] In this embodiment, by acquiring historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, sufficient and realistic basic data can be provided for training the initial photovoltaic output prediction model, ensuring the effectiveness of the initial photovoltaic output prediction model training. Next, by constructing the initial photovoltaic output prediction model, a mapping framework between photovoltaic output, irradiance, and temperature is established, providing a basic architecture for subsequent optimization of the initial photovoltaic output prediction model and ensuring the feasibility of the prediction logic. Then, using the historical irradiance dataset, historical temperature dataset, and historical photovoltaic dataset, the initial photovoltaic output prediction model is iteratively trained until a preset first convergence condition is met. This allows the initial photovoltaic output prediction model to fully learn the inherent correlation between photovoltaic output and irradiance and temperature in historical data, ultimately obtaining a preset photovoltaic output prediction model with satisfactory accuracy and strong stability, providing reliable support for subsequent accurate photovoltaic data prediction.

[0128] Furthermore, in the process of obtaining the predicted net load based on predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the construction process of the preset net load prediction model includes:

[0129] Obtain historical photovoltaic datasets, historical net load datasets, and historical projected net load datasets;

[0130] Construct an initial net load forecasting model;

[0131] Based on historical photovoltaic datasets, historical net load datasets, and historical predicted net load datasets, the initial net load prediction model is trained until the preset second convergence condition is met, thus obtaining the preset net load prediction model.

[0132] In this embodiment, by acquiring historical photovoltaic (PV) datasets, historical net load datasets, and historical predicted net load datasets, comprehensive and realistic basic data can be provided for training the initial net load prediction model, ensuring the effectiveness and relevance of the initial net load prediction model training. Next, an initial net load prediction model is constructed using a Long Short-Term Memory (LSTM) network, establishing a mapping framework between PV output, net load, and predicted net load, providing an architecture for subsequent optimization of the initial net load prediction model. Then, using the historical PV dataset, historical net load dataset, and historical predicted net load dataset, the initial net load prediction model is iteratively trained until a preset second convergence condition is met. This allows the initial net load prediction model to fully learn the nonlinear coupling influence of PV output on net load in historical data and the temporal variation characteristics of net load itself. Ultimately, a preset net load prediction model with satisfactory accuracy and strong stability is obtained, providing reliable support for subsequent accurate output of predicted net load and avoiding the accumulation of errors caused by independent predictions.

[0133] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for energy storage scheduling based on multi-source prediction, characterized in that, include: Acquire initial historical net load, measured photovoltaic data, real-time irradiance, initial real-time temperature, historical electricity price data, and real-time electricity price data; Based on the initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model, and preset net load prediction model, the predicted photovoltaic data, predicted net load, and predicted electricity price data are obtained, and a multi-source prediction sequence is constructed based on the predicted photovoltaic data, predicted net load, and predicted electricity price data. The real-time energy storage SOC trajectory is obtained, and a battery physical lifetime loss term and a grid interaction stability term are constructed based on the real-time energy storage SOC trajectory, predicted photovoltaic data and preset scheduling load. A multi-objective function is then constructed based on the grid interaction stability term and the battery physical lifetime loss term. Based on the initial historical net load, measured photovoltaic data, predicted photovoltaic data, and predicted net load, a probability density function for photovoltaic load prediction error is established, and several sets of simulated photovoltaic load curves are generated based on the probability density function for photovoltaic load prediction error. With the goal of minimizing a multi-objective function, several sets of mathematical expectation values ​​for simulated photovoltaic (PV) load curves are calculated based on these curves. Then, based on preset power balance constraints, preset battery state of charge constraints, preset mathematical expectation conditions, and these mathematical expectation values, the simulated PV load curves are filtered to obtain a first PV load curve. Finally, based on this first PV load curve, energy storage output data and the State of Charge (SOC) activation threshold are determined. The mathematical expectation value of the simulated PV load curve can be calculated using the following formula: ; Let S be the s-th simulated photovoltaic load curve, where S is the total number of simulated photovoltaic load curves. Let be the multi-objective function value of the s-th simulated photovoltaic load curve; The multi-source prediction sequence, energy storage output data, and SOC action threshold are sent to the energy storage converter so that the energy storage converter can complete the energy storage scheduling of the power system based on the multi-source prediction sequence, energy storage output data, and SOC action threshold.

2. The energy storage scheduling method based on multi-source prediction according to claim 1, characterized in that, The process involves acquiring predicted photovoltaic data, predicted net load, and predicted electricity price data based on initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, a preset photovoltaic output prediction model, a preset electricity price prediction model, and a preset net load prediction model. A multi-source prediction sequence is then constructed based on these data, including: Data cleaning is performed on the initial historical net load, real-time irradiance, and initial real-time temperature to obtain the first historical net load, moving average irradiance, and first real-time temperature. Based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, predictive photovoltaic data is obtained. Based on the predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the predicted net load is obtained. Based on historical electricity price data, real-time electricity price data, and a preset electricity price prediction model, predictive electricity price data is obtained. A multi-source prediction sequence is constructed based on predicted photovoltaic data, predicted net load, and predicted electricity price data.

3. The energy storage scheduling method based on multi-source prediction according to claim 2, characterized in that, In the process of acquiring predicted photovoltaic data based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, the construction process of the preset photovoltaic output prediction model includes: Obtain historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets; Construct an initial photovoltaic power output prediction model; Based on historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, the initial photovoltaic output prediction model is trained until the preset first convergence condition is met, thus obtaining the preset photovoltaic output prediction model.

4. The energy storage scheduling method based on multi-source prediction according to claim 2, characterized in that, In the process of obtaining the predicted net load based on predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the construction process of the preset net load prediction model includes: Obtain historical photovoltaic datasets, historical net load datasets, and historical projected net load datasets; Construct an initial net load forecasting model; Based on historical photovoltaic datasets, historical net load datasets, and historical predicted net load datasets, the initial net load prediction model is trained until the preset second convergence condition is met, thus obtaining the preset net load prediction model.

5. The energy storage scheduling method based on multi-source prediction according to claim 1, characterized in that, The process involves acquiring the real-time energy storage SOC trajectory, constructing a battery physical lifetime loss term and a grid interaction stability term based on the real-time energy storage SOC trajectory, predicted photovoltaic data, and preset dispatch load, and then constructing a multi-objective function based on the grid interaction stability term and the battery physical lifetime loss term, including: The real-time energy storage SOC trajectory is obtained, and the discharge depth and cycle temperature of several charge-discharge cycles are extracted based on the real-time energy storage SOC trajectory. A battery physical lifetime loss term is constructed based on depth of discharge and cycle temperature; A grid interaction stability term is constructed based on predicted photovoltaic data and preset dispatch loads. A multi-objective function is constructed based on the grid interaction stability term and the battery physical lifetime loss term.

6. An energy storage dispatch system based on multi-source prediction, characterized in that, It includes a multi-source data acquisition module, a multi-source sequence construction module, a multi-objective function construction module, a scheduling parameter solving module, and a scheduling execution module, specifically: The multi-source data acquisition module is used to acquire initial historical net load, measured photovoltaic data, real-time irradiance, initial real-time temperature, historical electricity price data, and real-time electricity price data; The multi-source sequence construction module is used to acquire predicted photovoltaic data, predicted net load and predicted electricity price data based on the initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model and preset net load prediction model, and to construct a multi-source prediction sequence based on the predicted photovoltaic data, predicted net load and predicted electricity price data; The multi-objective function construction module is used to obtain the real-time energy storage SOC trajectory, and construct a battery physical lifetime loss term and a grid interaction stability term based on the real-time energy storage SOC trajectory, predicted photovoltaic data and preset scheduling load, so as to construct a multi-objective function based on the grid interaction stability term and the battery physical lifetime loss term. The scheduling parameter solving module is used to establish a photovoltaic load prediction error probability density function based on the initial historical net load, measured photovoltaic data, predicted photovoltaic data and predicted net load, and generate several sets of simulated photovoltaic load curves based on the photovoltaic load prediction error probability density function. The scheduling parameter solving module aims to minimize a multi-objective function. Based on several sets of simulated photovoltaic load curves, it calculates several mathematical expectation values ​​for these curves. Then, based on preset power balance constraints, preset battery state of charge constraints, preset mathematical expectation conditions, and these mathematical expectation values, it filters the simulated photovoltaic load curves to obtain a first photovoltaic load curve. Finally, based on this first photovoltaic load curve, it determines the energy storage output data and the SOC (State of Charge) action threshold. The mathematical expectation value of the simulated photovoltaic load curve can be calculated using the following formula: ; Let S be the s-th simulated photovoltaic load curve, where S is the total number of simulated photovoltaic load curves. Let be the multi-objective function value of the s-th simulated photovoltaic load curve; The scheduling execution module is used to send the multi-source prediction sequence, energy storage output data and SOC action threshold to the energy storage converter, so that the energy storage converter can complete the energy storage scheduling of the power system based on the multi-source prediction sequence, energy storage output data and SOC action threshold.

7. The energy storage dispatch system based on multi-source prediction according to claim 6, characterized in that, The multi-source sequence construction module is used to acquire predicted photovoltaic data, predicted net load, and predicted electricity price data based on initial historical net load, real-time irradiance, initial real-time temperature, historical electricity price data, real-time electricity price data, preset photovoltaic output prediction model, preset electricity price prediction model, and preset net load prediction model, and to construct a multi-source prediction sequence based on the predicted photovoltaic data, predicted net load, and predicted electricity price data, including: Data cleaning is performed on the initial historical net load, real-time irradiance, and initial real-time temperature to obtain the first historical net load, moving average irradiance, and first real-time temperature. Based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, predictive photovoltaic data is obtained. Based on the predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the predicted net load is obtained. Based on historical electricity price data, real-time electricity price data, and a preset electricity price prediction model, predictive electricity price data is obtained. A multi-source prediction sequence is constructed based on predicted photovoltaic data, predicted net load, and predicted electricity price data.

8. The energy storage dispatch system based on multi-source prediction according to claim 7, characterized in that, In the process of acquiring predicted photovoltaic data based on the moving average irradiance, the first real-time temperature, and the preset photovoltaic output prediction model, the construction process of the preset photovoltaic output prediction model includes: Obtain historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets; Construct an initial photovoltaic power output prediction model; Based on historical irradiance datasets, historical temperature datasets, and historical photovoltaic datasets, the initial photovoltaic output prediction model is trained until the preset first convergence condition is met, thus obtaining the preset photovoltaic output prediction model.

9. The energy storage dispatch system based on multi-source prediction according to claim 7, characterized in that, In the process of obtaining the predicted net load based on predicted photovoltaic data, the first historical net load, and the preset net load prediction model, the construction process of the preset net load prediction model includes: Obtain historical photovoltaic datasets, historical net load datasets, and historical projected net load datasets; Construct an initial net load forecasting model; Based on historical photovoltaic datasets, historical net load datasets, and historical predicted net load datasets, the initial net load prediction model is trained until the preset second convergence condition is met, thus obtaining the preset net load prediction model.

Images