A high-efficiency energy scheduling and intelligent control method for a photovoltaic energy storage system

Abstract

This invention relates to the field of photovoltaic energy storage, specifically to a method for efficient energy dispatch and intelligent control of photovoltaic energy storage systems. The method comprises the following steps: S1: Constructing a multi-source data acquisition platform to collect and preprocess photovoltaic power generation data, energy storage device data, load demand data, and meteorological data; S2: Based on the preprocessed data in S1, using the Gradient Boosting Decision Tree (GBDT) algorithm, constructing a photovoltaic power generation prediction model and a load demand prediction model; S3: Based on the output of the prediction models, using a model predictive control algorithm, generating the optimal control sequence in the rolling time domain. The advantages of this invention are: In the research of photovoltaic energy storage systems, a method for efficient energy dispatch and intelligent control based on big data is proposed, aiming to improve the overall operating efficiency of photovoltaic power generation and energy storage systems, optimize power dispatch, and achieve intelligent energy storage control.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic energy storage, and specifically to a method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system. Background Technology

[0002] With the global energy structure transformation and the rapid development of renewable energy technologies, photovoltaic (PV) power generation, as an important component of clean energy, is increasingly accounting for a larger share of energy supply. However, the intermittency and volatility of PV power generation pose significant challenges to the stable operation of the power grid. To effectively address this issue, PV energy storage systems have emerged as a key technology for mitigating power fluctuations and improving system reliability. During the operation of PV energy storage systems, energy dispatch and intelligent control are crucial for ensuring efficient system operation. Traditional energy dispatch methods rely primarily on simple rule models or empirical judgments, which are ill-suited to the complex and ever-changing scenarios of PV power generation and load demand. Furthermore, the control strategies of energy storage devices are often overly conservative or aggressive, leading to shortened equipment lifespan or reduced system economics. The rapid development of big data and artificial intelligence technologies provides new technical support for the energy dispatch and intelligent control of PV energy storage systems. By constructing a big data platform, comprehensive data collection and analysis of PV power generation, energy storage devices, load demand, and meteorological data can be achieved, providing a data foundation for system operation decisions. However, effectively utilizing massive amounts of data to construct accurate predictive models and implement intelligent control strategies remains a key research focus and challenge. Furthermore, as a crucial component of the smart grid, photovoltaic energy storage systems need to work collaboratively with the power grid. Optimizing the system's power supply and energy storage strategies based on peak and off-peak electricity prices and load conditions to improve economic efficiency and energy utilization is a critical issue that urgently needs to be addressed. Therefore, developing a data-driven, efficient energy dispatch and intelligent control method for photovoltaic energy storage systems is of significant theoretical and practical importance for improving overall system operating efficiency and promoting the development of the new energy industry.

[0003] Currently, energy dispatching for photovoltaic energy storage systems mainly employs rule-based dispatching and dynamic programming methods. These methods have limitations when dealing with complex dynamic environments and are difficult to adapt to rapid changes in photovoltaic power generation and load demand. Existing photovoltaic energy storage system energy dispatching technologies mainly include rule-based dispatching algorithms, dynamic programming algorithms, and heuristic algorithms. These methods have the following limitations in practical applications: (1) Rule-based dispatching algorithms are too simple and difficult to adapt to complex and ever-changing photovoltaic power generation and load demand scenarios, resulting in low dispatching efficiency; (2) Dynamic programming algorithms have high computational complexity and are difficult to meet the real-time dispatching requirements of large-scale photovoltaic energy storage systems; (3) Heuristic algorithms, such as genetic algorithms and particle swarm optimization algorithms, are prone to getting trapped in local optima and are difficult to guarantee global optimization effects.

[0004] Traditional energy storage control strategies are mostly based on fixed thresholds or simple charge and discharge rules, which cannot fully consider the dynamic characteristics of energy storage devices and the overall operating status of the system, resulting in low utilization efficiency and shortened lifespan of energy storage devices. At present, the control technologies for energy storage devices mainly include empirical rule-based control methods and model-based predictive control methods. These methods have the following shortcomings: (1) Empirical rule-based control methods lack an accurate description of the dynamic characteristics of energy storage devices, making it difficult to achieve precise control; (2) Traditional model predictive control methods are highly dependent on the system model, making it difficult to adapt to the nonlinear and time-varying characteristics of photovoltaic energy storage systems; (3) Existing control methods do not fully consider the lifespan characteristics of energy storage devices, making it difficult to achieve long-term optimized operation of energy storage devices.

[0005] In recent years, big data technology has been increasingly widely used in the field of energy management. However, existing research focuses on the optimization of single links and lacks unified modeling and collaborative optimization of photovoltaic power generation, energy storage control and load demand. In recent years, research on big data-driven energy management systems has made some progress, but the following problems still exist: (1) Existing data-driven methods are mostly single algorithms, which are difficult to balance prediction accuracy and computational efficiency at the same time; (2) There is a lack of effective integration and utilization of multi-source heterogeneous data, making it difficult to fully explore the value of data; (3) Existing methods have failed to effectively combine photovoltaic power generation prediction, load prediction and energy storage control, making it difficult to achieve system-level collaborative optimization.

[0006] In summary, the above analysis demonstrates that big data-driven, efficient energy dispatch and intelligent control methods for photovoltaic energy storage systems are crucial for improving renewable energy utilization and optimizing power system operation. The GBDT algorithm effectively processes multi-source heterogeneous data, improving the accuracy of photovoltaic power generation and load demand forecasting, and providing a reliable basis for energy dispatch decisions. The MPC algorithm, by dynamically optimizing the charging and discharging process of energy storage devices, can extend equipment lifespan and reduce operating costs. Combining GBDT and MPC algorithms in photovoltaic energy storage systems helps achieve intelligent dispatch and control of the system, improving overall operational efficiency.

[0007] Therefore, this invention aims to develop a big data-driven, high-efficiency energy dispatch and intelligent control method for photovoltaic energy storage systems to solve problems such as inefficient energy dispatch and insufficient energy storage control in existing technologies. Specific objectives include: (1) constructing a GBDT-based photovoltaic power generation and load demand prediction model to improve prediction accuracy; (2) designing an MPC-based intelligent control algorithm for energy storage devices to optimize charging and discharging strategies; and (3) achieving intelligent collaborative operation between photovoltaic power generation, energy storage systems, and the power grid to improve system economic benefits and energy efficiency. This invention is expected to provide an innovative solution for the integration and application of large-scale photovoltaic power generation and energy storage systems, promoting the efficient utilization of renewable energy and the development of smart grids. Summary of the Invention

[0008] To address the problems of inefficient energy dispatch and insufficient energy storage control in existing technologies, this invention provides a method for efficient energy dispatch and intelligent control of photovoltaic energy storage systems, as detailed below:

[0009] A method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system, the energy dispatch process of which includes the following steps: S1: Construct a multi-source data acquisition platform to collect and preprocess data including photovoltaic power generation data, energy storage equipment data, load demand data, and meteorological data. S2: Based on the preprocessed data in S1, the gradient boosting decision tree algorithm, i.e., the GBDT algorithm, is used to construct a photovoltaic power generation prediction model and a load demand prediction model. S3: Based on the output of the prediction model, a model predictive control algorithm, namely the MPC algorithm, is used for predictive control. The goal is to minimize the comprehensive cost including operating costs and energy storage losses. The charging and discharging strategy of the energy storage device is optimized in a rolling manner to generate the optimal control sequence in the rolling time domain. S4: Based on real-time grid time-of-use pricing, construct a grid-energy storage collaborative optimization model with the goal of minimizing grid interaction costs, and formulate a power purchase and sale strategy with the grid; S5: Based on system operation feedback, adaptively update the parameters of the two prediction models and control strategies in step S2.

[0010] The multi-source data acquisition platform includes: a photovoltaic array monitoring module for collecting string-level current, voltage, and module backsheet temperature data; an energy storage system data acquisition module for acquiring real-time battery state of charge (SOC), state of health (SOH), and temperature distribution information; a meteorological sensing unit for simultaneously collecting meteorological data such as solar irradiance, ambient temperature, and cloud cover; and a load monitoring module for acquiring time-of-use electricity load curve data through smart meters.

[0011] Furthermore, in step S2, a photovoltaic power generation prediction model and a load demand prediction model are constructed. The photovoltaic power generation prediction models are all based on the GBDT algorithm and are collectively referred to as GBDT energy prediction models. The specific construction steps are as follows: (S2.1) Integrate historical photovoltaic power generation data, meteorological characteristics, and load time series data to construct a multi-dimensional input feature set; (S2.2) Generate training label data based on the input features and construct a nonlinear regression model; (S2.3) Determine the optimal parameter combination for the decision tree through cross-validation and hyperparameter optimization; (S2.4) An iterative correction mechanism is adopted to generate weak learners in each round and optimize and correct the prediction error of the previous model until the convergence condition is met.

[0012] Historical photovoltaic (PV) power generation data is used to capture autocorrelation and periodic patterns. For example, yesterday's power generation at the same time can be used to predict today's power generation at the same time. Training labels are generated based on input features, and a nonlinear regression model is constructed. The goal is to define the prediction task and initialize the model, clarifying what the model will "teach" and selecting GBDT as the "learning framework." The generated training labels are typically the PV active power at a future point in time. Because the prediction is a continuous value (power value), it is a regression task, not a classification task. Nonlinearity: GBDT itself is a nonlinear model. Here, "construction" mainly refers to determining the model's loss function. For regression problems, mean squared error is commonly used as the loss function; that is, the model's goal is to minimize the sum of the squares of the differences between the predicted and true values.

[0013] By cross-validation and hyperparameter optimization, the optimal parameter combination is determined. The goal is to find the parameter settings that make the model have the strongest generalization ability and are least prone to overfitting, and to tune the GBDT "engine" to the best performance.

[0014] Hyperparameter optimization refers to using grid search or random search to find the best-performing set of hyperparameters in a predefined space of combinations, aiming to balance bias and variance. Trees that are too deep or numerous may perfectly fit the training data (low bias), but will predict poorly when encountering new data (high variance, i.e., overfitting). Optimization finds the point that performs well on both the training and validation sets.

[0015] The iterative correction mechanism, which generates a weak learner and optimizes the error, is the core process for actually training the GBDT model. The first iteration calculates the current model's prediction residual as the true label minus the current predicted value. Then, a new decision tree is trained. However, this tree's goal is not to predict the original label, but to learn and fit the residual calculated in the previous step. This tree is a "weak learner." The prediction result of this tree (multiplied by a small learning rate) is added to the initial prediction, forming a corrected prediction. At this point, the prediction result is closer to the true value than the initial value. The Nth iteration then proceeds: the new residual predicted after the previous correction is calculated, and a new tree is trained to fit this new residual. The prediction result of the new tree (weighted) is added to the prediction result of the previous round for further correction.

[0016] This process is repeated continuously. Each new tree strives to correct the errors accumulated by all the previous trees, until the preset convergence condition is met.

[0017] Furthermore, in step S3, the MPC algorithm is used for predictive control, specifically including: (S3.1) First, define the system's state variables and control variables. Let the system's state vector be... It includes at least the photovoltaic power generation capacity, the charge / discharge state of the energy storage device, and the load demand, with the control vector being... If the system includes at least the charging and discharging power of the energy storage device and the load distribution, then the dynamic model of the system can be expressed as: (Formula 1) in, It is the system's state transition function. This represents random perturbations in the system, reflecting the uncertainty of photovoltaic power generation. (S3.2) Next, define the control objective and constraints: the objective function can be expressed as: (Formula 2) in, To predict the length of the time domain, and These are the costs of electricity purchase and energy storage losses, respectively. and The first The power purchase capacity and energy storage capacity at any given time.

[0018] (S3.3) A rolling optimization and online update mechanism is adopted, with the prediction time domain set to 24 hours and the control time domain set to 1 hour.

[0019] In this algorithm, the system's state variables and control variables are defined, and the system's state vector is assumed to be... This includes photovoltaic power generation, the charging and discharging status of energy storage devices, and load demand; the control vector is... This includes the charging and discharging power of the energy storage device and load distribution; the system refers to the photovoltaic energy storage system, including the overall operating environment of photovoltaic power generation equipment, energy storage equipment, and load equipment; state variables. Describe the system's operating state at time t, including dynamic characteristics such as photovoltaic power generation, energy storage device charging and discharging state, and load power demand; control variables Describe the system's operational inputs, including the charging and discharging power of the energy storage devices, load allocation, and power scheduling strategies;

[0020] The control objective is to optimize the overall energy efficiency and economy of the system by minimizing the total system operating cost over the future forecast period. This objective function describes the total cost of the control system over the rolling forecast control period, specifically including electricity purchase costs and energy storage loss costs.

[0021] The constraints include physical and operational constraints of the system, including that the charging and discharging power of the energy storage device should meet its rated power limit, the state of the energy storage device should be within its capacity range, and the load requirements should be met.

[0022] To improve the real-time performance and robustness of the algorithm, rolling optimization and online update mechanisms are introduced. Through rolling optimization, the control strategy is implemented only at the current moment in each control cycle and re-optimized in the next cycle to dynamically adapt to changes in the system. Through online updates, the latest information on photovoltaic power generation and load demand is obtained in real time to update the system model and predictions, thereby further improving the accuracy and effectiveness of the control strategy.

[0023] Furthermore, in step S4, constructing the grid-energy storage collaborative optimization model includes the following steps: (1) Construct a time-of-use pricing response function to characterize the economics of the interaction between the system and the power grid. The response function is as follows: (Formula 3) in, The power grid interaction cost function at time t; This indicates the power purchased from the power grid. This indicates the power output sold to the grid. and These represent the unit price of electricity purchase and sales or the electricity price weighting factor for the corresponding time period.

[0024] (2) Design the priority rules for energy storage charging and discharging: prioritize charging during off-peak hours and prioritize discharging during peak hours; through the time-of-use pricing mechanism, minimize the purchase of electricity from the grid during peak hours and sell electricity to the grid using the energy storage system during off-peak hours, thereby improving the grid interaction economy of the system.

[0025] (3) Set grid power interaction constraints: To meet the requirements of system operation safety and power purchase capacity constraints.

[0026] (4) Construct an optimization model with the goal of minimizing grid interaction costs, and use dynamic programming algorithm to solve for the optimal power purchase and sale strategy within the daily cycle range: (Formula 4)

[0027] By constructing a time-of-use (TOU) electricity price response function to characterize the economics of the system's interaction with the grid, a charging / discharging priority rule is designed. Charging operations are performed during off-peak hours when electricity prices are low, and discharging is planned during peak hours. The core of the rule is: "buying" when electricity prices are lowest and "selling" when prices are highest, profiting from the price difference. This is the most direct economic driver, resulting in a significant reduction in overall electricity costs: replacing high-priced grid electricity with self-generated power drastically reduces electricity bills. Indirect social benefits include peak shaving and valley filling: actively adjusting the electricity load curve. Discharging during peak hours reduces the supply pressure on the grid, avoiding the investment in expensive generation and transmission facilities to meet short-term peak demand. Charging during off-peak hours increases the stable load on the grid and improves the utilization efficiency of power generation equipment.

[0028] Energy storage enhances grid stability and reliability, providing rapid response, backup capacity, and smoothing out renewable energy fluctuations. It also allows users to profit from seasonal price differences while mitigating market volatility.

[0029] Furthermore, in constructing the GBDT energy forecasting model, it is assumed that photovoltaic power generation and load demand are denoted as follows: and Then, the following formula can be used for prediction: (Formula 5) (Formula 6) in, and These are the weights of the decision tree. and For the output of the decision tree, and The number of decision trees.

[0030] In this method, the parallel computing capabilities of MPC are utilized to dynamically optimize the energy prediction model and load demand prediction model using a parallelized processor architecture, thereby achieving a real-time and efficient energy allocation strategy. Furthermore, the combination of MPC with the GBDT algorithm significantly improves both prediction accuracy and optimization efficiency.

[0031] Furthermore, it also includes the calculation process of minimizing the prediction error of the GBDT energy prediction model, including finding a function By progressively optimizing the loss function, the prediction error is minimized. This involves finding the function... The method is as follows: (1) Initialize the model: (Formula 5) in, Represents the loss function. γ represents the target variable, γ represents the initial values ​​of the model parameters, and N represents...? (3) For each iteration Calculate the residuals using the following formula: (Formula 6) Where: m represents the current iteration round, increasing sequentially from 1 to M, and M is the total number of iteration rounds; r im Let F(x) represent the residual of the i-th data sample in the m-th round, which is used to measure the deviation between the current model prediction and the target value; F(x) is the output prediction value of the model in the m-th round, which is updated based on the model results of the previous round. F represents the target variable. m-1 (x) represents the model's prediction output function in the m-1th round.

[0032] (4) Fit a new decision tree This allows it to minimize the sum of squares of the residuals: (Formula 7) (5) Update the model: (Formula 8) in, This represents the output of the prediction model after the m-th iteration; This represents the model output from the previous iteration; In the first The weak learner (regression tree) obtained during training is used to fit the residuals. ; Let be the learning rate, where 0 < 1. ≤1 is used to control the contribution ratio of each tree to the final model.

[0033] Through the above iterative update process, the energy prediction model based on GBDT can gradually optimize the loss function. Ultimately, a model capable of accurately predicting photovoltaic power generation energy was obtained. .

[0034] We construct weak learners, i.e., decision trees, using multi-scale pulse-coupled gradient boosting. Each tree in the decision tree corrects the prediction error of the previous tree. Here, we assume a training dataset. ,in Indicates input features, Let L represent the target variable; N represents the total number of samples in the training dataset, i.e., the number of feature-target pairs; i represents the sample number in the training data, which is then used to initialize the model. Prediction error refers to the deviation between the model's predicted value and the actual observed value. Commonly used loss functions include mean squared error (MSE) and mean absolute error (MAE). γ in Equation 5 represents the initial value of the model parameters, used to determine the optimal parameters by minimizing the loss function L during model initialization, thus providing initial parameter configuration for subsequent optimization processes.

[0035] Through the above iterative update process, the energy prediction model based on GBDT can gradually optimize the loss function. Ultimately, a model capable of accurately predicting photovoltaic power generation energy was obtained. The above method, which involves constructing and progressively optimizing a loss function to minimize this error, is known as "finding a prediction function". The process has the following characteristics: aiming to minimize prediction error; training a new model based on residuals in each round; the model gradually converges after multiple iterations; learning rate; It determines the convergence speed and the model's generalization ability.

[0036] Furthermore, step S3 involves rolling optimization of the energy storage device's charging and discharging strategy, based on the MPC control algorithm, including formulating an energy storage discharge optimization strategy, assuming the energy storage device's charging power is... The discharge power is The optimization objective is then expressed as: (Formula 9) in, and The cost functions for charging and discharging are respectively. To optimize the time period, and These represent the charging and discharging power of the energy storage system at time t, respectively.

[0037] An energy storage discharge optimization strategy is formulated, which takes the charging and discharging cost of energy storage devices as the optimization objective. By building a predictive model, the charging and discharging power allocation of the energy storage system is dynamically adjusted according to real-time power demand and energy storage status, so as to minimize operating costs and maximize the lifespan of energy storage devices. By dynamically adjusting the charging and discharging power in real time, the optimization strategy reduces system operating costs while meeting power demand.

[0038] Furthermore, it also includes an evaluation of the performance of the GBDT-based energy prediction model, the evaluation method including cross-validation or hyperparameter optimization, wherein the cross-validation method includes the following steps: (1) Dataset definition: The dataset is a complete dataset composed of multi-source data such as photovoltaic power generation systems, energy storage devices, and load demand, denoted as , where x i Represents the input features, y i Represent the target variable; (2) Cross-validation method: Divide the dataset into k subsets. For each fold i=1,2,....,k, select the i-th subset as the validation set and the remaining k-1 subsets as the training set. Train the GBDT model on the training set each time and evaluate the model performance on the validation set. Record the model performance index. After cross-validation is completed, calculate the average value of the validation results of each fold as the overall evaluation standard of the model performance. (3) Performance evaluation: After cross-validation is completed, the average value of all K rounds of validation results is calculated as the overall performance index of the model;

[0039] The hyperparameter optimization method is as follows: before model training, the optimal combination of hyperparameters is searched to improve the performance of the model.

[0040] The hyperparameter optimization methods include grid search and random search; the optimized hyperparameters include the maximum depth of the decision tree, the learning rate, and the subsample ratio.

[0041] After updating the model, the GBDT model iteratively optimizes the loss function, ultimately yielding a model capable of accurately predicting photovoltaic power generation. To further enhance the model's robustness and generalization ability, this invention introduces cross-validation and hyperparameter optimization techniques during model construction. Cross-validation is a technique for evaluating model performance. It involves dividing the dataset into multiple mutually exclusive subsets, using one subset sequentially as the validation set and the remaining subsets as the training set, and finally evaluating the model's performance by averaging the results of multiple validations.

[0042] Furthermore, the constraints include physical and operational constraints of the system, including that the charging and discharging power of the energy storage device should meet its rated power limit, the state of the energy storage device should be within its capacity range, and the load demand should be met. The above three constraints are expressed as follows: Formula 10: (Formula 10) in, and The minimum and maximum charge and discharge power of the energy storage device, and The minimum and maximum energy states of an energy storage device. For the first The load power requirement at any given time.

[0043] Within each control cycle, based on the current system state By solving the above optimization problem, the optimal control sequence for a future period of time can be obtained. Then, apply the control strategy for the current moment. And repeat the above process in the next control cycle.

[0044] Furthermore, it also includes predicting photovoltaic power generation. The GBDT energy prediction model uses historical data and environmental factors to predict future photovoltaic power generation. Let the predicted photovoltaic power generation be... ,but: (Formula 11)

[0045] Furthermore, intelligent scheduling is achieved by combining the MPC algorithm. The MPC algorithm, by integrating historical data and environmental factors, predicts grid demand and photovoltaic power generation over a future period, thereby optimizing the charging and discharging strategies of energy storage devices. Specifically, formulas 12 and 13 describe the optimization objectives for power balance and scheduling. Formula 14 dynamically models the state update of energy storage devices based on their characteristics, realizing dynamic energy scheduling based on the innovative points of this invention. Let the predicted grid demand be... The MPC optimization problem can then be described as follows: (Formula 12) Of which, photovoltaic power generation is The charging power of the energy storage device is The discharge power is The demand of the power grid is The charging and discharging efficiencies of the energy storage devices are respectively and ; The power grid demand is The equilibrium equation can be expressed as: (Formula 13) The state update equation for the energy storage device is: (Formula 14) Among them, the status of energy storage devices is Its capacity is ; Constraints: (Formula 15) (Formula 16) (Formula 17)

[0046] Formulas 15 and 16 define the dynamic range of the charging and discharging power of energy storage devices, while Formula 17 further considers the state of the energy storage device. and capacity To achieve real-time control of energy storage capacity and ensure the safety and efficiency of system operation, these constraints are combined with the modeling framework of existing technologies, and the optimized design of this invention is formed by incorporating dynamic scheduling objectives and innovative state update equations.

[0047] To predict photovoltaic power generation The GBDT model is introduced. The GBDT model uses historical data and environmental factors (such as solar irradiance and temperature) to predict future photovoltaic power generation.

[0048] When formulating a scheduling strategy, it also includes intelligent coordination with the power grid, combining peak-valley electricity pricing strategies and power load status to achieve intelligent scheduling of power supply and energy storage.

[0049] The advantages of this invention are:

[0050] In the research of photovoltaic energy storage systems, this invention proposes a high-efficiency energy dispatch and intelligent control method based on big data, aiming to improve the overall operating efficiency of photovoltaic power generation and energy storage systems, optimize power dispatch, and realize intelligent energy storage control. Specifically, it includes: (1) constructing a unified prediction model of photovoltaic power generation and load demand based on the GBDT algorithm, which significantly improves the prediction accuracy and provides reliable data support for subsequent energy dispatch; (2) using the MPC algorithm to dynamically adjust the charging and discharging plan of energy storage devices through a rolling optimization strategy, effectively balancing the immediate demand and long-term benefits of the system. This method effectively overcomes the shortcomings of traditional rule-based control algorithms in dealing with the dynamic changes of complex systems, significantly extends the service life of energy storage devices, and reduces the operating cost of the system; (3) proposing an intelligent collaborative dispatch strategy to realize the coordinated operation of photovoltaic power generation, energy storage systems, and the power grid. This strategy maximizes the economic benefits and energy utilization rate of the system. (4) Through simulation experiments, this invention verifies that the proposed method can improve the overall energy utilization rate of the system by about 15% and reduce grid fluctuations by about 20% compared with traditional scheduling algorithms; (5) An adaptive data processing and model update mechanism was developed, which significantly improves the system's adaptability to environmental changes and equipment aging, ensuring long-term operational stability and efficiency. Experimental results show that this mechanism reduces the fluctuation of scheduling efficiency of the system under different seasons and weather conditions by about 30%. Attached Figure Description

[0051] Figure 1 This is a flowchart of the present invention;

[0052] Figure 2 This is a comparison of energy dispatch efficiency in Example 1;

[0053] Figure 3 Example 1: Comparison of photovoltaic power generation prediction accuracy;

[0054] Figure 4 The charging and discharging optimization effect of the energy storage system in Example 1;

[0055] Figure 5 Example 1 illustrates the improvement in system energy utilization under different weather conditions;

[0056] Figure 6 This is a comparison of energy scheduling execution time in Example 1. Detailed Implementation

[0057] Example 1

[0058] The efficient energy dispatch and intelligent control method of photovoltaic energy storage system was simulated and tested using the MATLAB platform to verify the effectiveness of the control strategy and optimization algorithm. The simulation model includes the main components of data acquisition and preprocessing module, GBDT algorithm module and MPC algorithm module, which can realistically reflect the operating characteristics and control behavior of the optimization control method applied to photovoltaic energy storage system.

[0059] The simulation model was subjected to various scenarios and reworked, including different weather conditions, load changes, and fluctuations in photovoltaic power generation, to test the system's adaptability and robustness. The specific hardware configuration is as follows: Processor: Intel Xeon Gold 6248 (2.5 GHz); Memory: 256 GB DDR4 RAM; Storage: 1 TB NVMe SSD; The software environment configuration is as follows: Operating System: Windows Server 2019; Simulation Platform: MATLAB R2021a / Simulink; Data Processing Tool: Python 3.8 for data preprocessing and analysis; Database: MySQL 8.0 for storing and managing big data.

[0060] (1) Setting simulation parameters:

[0061] The simulation model's parameter settings referenced the technical specifications and operational data of the optimized control method applied to photovoltaic energy storage systems, ensuring the simulation results have practical significance. The settings of key simulation parameters are shown in Table 1. The dataset for the optimized control method applied to photovoltaic energy storage systems was set as a typical national photovoltaic power generation dataset, covering the period from 2019 to 2021. The specific parameter settings of the simulation model are shown in Table 1.

[0062] Table 1 Parameter Settings

[0063]

[0064] (2) Results analysis is shown in the comparison of energy dispatch efficiency. Figure 2 ;

[0065] Figure 2 The comparison results of four different algorithms in terms of energy dispatch efficiency are presented. First, under sunny conditions, the dispatch efficiency of each algorithm is generally good. The MPC algorithm has the highest efficiency at 94.3%, with a small error range, indicating that it can effectively optimize energy dispatch under stable weather conditions, demonstrating high robustness and stability. The GBDT algorithm follows closely with a dispatch efficiency of 92.5%, although its error is slightly larger than that of the MPC algorithm, its overall performance is still excellent. In contrast, the efficiency of the traditional rule-based algorithm is significantly lower, at only 78.2%, and its error range is larger, reflecting its insufficient flexibility in dealing with photovoltaic power generation system dispatch and its inability to effectively handle complex dispatch tasks. The SVR algorithm has an efficiency of 85.8%, showing moderate dispatch efficiency, indicating that it can work effectively under static conditions, but its dispatch optimization capability is not as good as the GBDT and MPC algorithms. Under cloudy conditions, as weather volatility increases, the dispatch efficiency of all algorithms decreases. The MPC algorithm still shows the highest dispatch efficiency at 92.8%, with a small error range, further proving that this algorithm can effectively optimize energy allocation when dealing with dynamic changes. The efficiency of the GBDT algorithm dropped to 90.1%, with a slight increase in the error range, but it still demonstrated high adaptability. The traditional rule-based algorithm performed poorly under cloudy conditions, with a scheduling efficiency of 75.6% and increased error, indicating its difficulty in flexibly adjusting strategies in the face of complex weather changes. The SVR algorithm also saw a decline in performance, with a scheduling efficiency of 83.2%, showing that it could not provide performance comparable to the MPC and GBDT algorithms under large weather fluctuations. Under rainy conditions, the complexity of energy scheduling further increased due to a significant reduction in photovoltaic power generation, and the scheduling efficiency of all algorithms continued to decline. The MPC algorithm maintained the highest efficiency at 91.5%, with a small error range, demonstrating its strong adaptability and robustness under extreme weather conditions. The efficiency of the GBDT algorithm dropped to 88.7%, but despite the decline, it still maintained high scheduling efficiency, indicating good dynamic response capabilities. The efficiency of the traditional rule-based algorithm dropped to 73.1%, with the largest error range, indicating significant instability and limitations under extreme weather conditions. The SVR algorithm performed slightly better than the traditional rule-based algorithm, but its scheduling efficiency was only 81.9%.

[0066] Figure 3This paper compares the prediction accuracy of three algorithms—GBDT, LSTM, and SVR—in terms of photovoltaic (PV) power generation. First, under clear weather conditions, the prediction errors of all three algorithms are relatively small and their distribution is relatively concentrated. The GBDT algorithm performs exceptionally well, with the smallest median prediction error and a short interquartile range in its box plot, indicating high prediction accuracy and low volatility under clear weather conditions, effectively adapting to stable PV power output. In contrast, the LSTM algorithm has a slightly larger prediction error and a relatively wider error distribution range, but it remains within an acceptable range. The SVR algorithm performs slightly worse, with a more dispersed prediction error distribution and a relatively large maximum prediction error, indicating that under stable weather conditions, the SVR algorithm cannot match the performance of GBDT and LSTM. Second, under cloudy weather conditions, the prediction errors of all algorithms increase. The uncertainty of weather poses a greater challenge to the fluctuation of PV power generation. The GBDT algorithm still maintains the smallest median prediction error and a narrow error range, demonstrating strong adaptability and the ability to make relatively accurate predictions under unstable power generation environments. The LSTM algorithm performed relatively stably, although its prediction error increased, but the distribution remained relatively concentrated. The SVR algorithm, however, showed a significant increase in prediction error, demonstrating insufficient adaptability to complex weather conditions. Its error range was significantly larger than the other two algorithms, and the maximum error point was significantly higher, indicating poor prediction accuracy under fluctuating weather conditions. Under cloudy conditions, due to a significant decrease in photovoltaic power generation, the prediction errors of all three algorithms reached their highest values. However, the GBDT algorithm maintained good performance, with its median prediction error and distribution range remaining optimal, indicating that it can effectively capture the characteristics of power generation fluctuations and provide relatively accurate prediction results under complex weather conditions. The LSTM algorithm's prediction error increased, but still maintained a relatively controllable error range, demonstrating good adaptability to power generation fluctuations under complex weather conditions. In contrast, the SVR algorithm's prediction error increased the most under cloudy conditions, with the most dispersed distribution, and the maximum error was significantly higher than the other two algorithms, reflecting the limitations of the SVR algorithm's prediction accuracy under extreme weather conditions.

[0067] Figure 4This paper compares the performance of three different algorithms in optimizing the charging and discharging of energy storage systems. First, under standard load conditions (Figure (a)), the charge / discharge depth curves of the three algorithms show significant differences. The MPC algorithm maintains a relatively smooth charge / discharge curve over 24 hours, especially during peak load periods (such as morning and afternoon peak hours), where its charge / discharge depth responds promptly to demand, maintaining around 95% with minimal overall fluctuation. This indicates that the MPC algorithm achieves dynamic scheduling of the system through predictive and optimized control, effectively improving the utilization efficiency of the energy storage system. In contrast, the rule-based control algorithm and the empirical rule-based scheduling algorithm show relatively low charge / discharge depths during peak periods, around 85%, with significant fluctuations, indicating that traditional algorithms have a lagging response during peak loads and struggle to achieve ideal control results. Second, under high load conditions (Figure (b)), the differences between the three algorithms become even more pronounced. The MPC algorithm still demonstrates strong load response capabilities, with its charge / discharge depth adjusting rapidly under high load conditions. Particularly during periods of high load, its charge / discharge depth remains close to 97%, ensuring energy scheduling needs are met under high load conditions. In contrast, the rule-based control algorithm and the empirical rule-based scheduling algorithm lagged behind, especially under high load conditions, where the charge and discharge depths remained at around 87% and 92%, respectively, failing to fully respond to the demands under high load. This indicates that traditional scheduling algorithms struggle to achieve optimal control of the energy storage system when faced with complex load variations. Finally, in low-load scenarios (Figure (c)), the overall charge and discharge depths of the three algorithms decreased compared to high load, and the curve fluctuations were relatively smaller. However, the MPC algorithm still demonstrated stable scheduling capabilities under low load conditions, maintaining a charge and discharge depth above 90% in all time periods. In contrast, the rule-based control algorithm and the empirical rule-based scheduling algorithm performed relatively weaker, with charge and discharge depths around 83% and 88%, respectively, and their scheduling flexibility and accuracy were still inferior to the MPC algorithm. This further demonstrates that the MPC algorithm possesses high adaptability and scheduling optimization capabilities under different load conditions.

[0068] Figure 5This paper presents a comparison of the GBDT+MPC algorithm with rule-based scheduling and dynamic programming algorithms in terms of system energy utilization. Firstly, under sunny conditions, all three algorithms demonstrate relatively ideal performance in photovoltaic (PV) power generation and energy storage utilization. The GBDT+MPC algorithm achieves the highest PV power generation utilization rate at 92.5%, and its energy storage utilization rate is also high at 88.3%. This high efficiency indicates that the GBDT+MPC algorithm can fully utilize the PV power generation system under sufficient sunlight through prediction and scheduling optimization, and effectively coordinate the charging and discharging process of the energy storage system, maximizing the overall energy utilization efficiency of the system. In contrast, the rule-based scheduling and dynamic programming algorithms perform slightly worse, with PV power generation and energy storage utilization rates of 78.4% and 84.7% (PV power generation), and 73.2% and 79.6% (energy storage), respectively. This suggests that traditional algorithms still suffer from some energy waste and insufficient scheduling efficiency when facing relatively stable power generation environments. Under cloudy conditions, the energy utilization rates of all three algorithms decrease, but the GBDT+MPC algorithm still maintains a significant advantage. The GBDT+MPC algorithm achieved a photovoltaic power generation utilization rate of 88.3% and an energy storage utilization rate of 85.7%, indicating that even under relatively unstable weather conditions, it can maintain a high system energy utilization rate through accurate prediction and flexible scheduling. In contrast, the photovoltaic power generation utilization rates of the rule-based scheduling algorithm and the dynamic programming algorithm were 75.2% and 81.6%, respectively, and the energy storage utilization rates were 70.5% and 76.8%, respectively. This demonstrates that traditional scheduling algorithms are less flexible in dealing with weather fluctuations and cannot fully optimize energy allocation, leading to a significant decrease in energy utilization. Under cloudy conditions, the energy utilization rates of all three algorithms further decreased due to a significant reduction in photovoltaic power generation. However, the GBDT+MPC algorithm still demonstrated strong adaptability under low light conditions, with photovoltaic power generation and energy storage utilization rates of 84.1% and 82.9%, respectively, significantly higher than the other two algorithms. In contrast, the photovoltaic power generation utilization rates of the rule-based scheduling algorithm and the dynamic programming algorithm were 71.8% and 77.9% respectively under cloudy conditions, while the energy storage utilization rates were 68.1% and 74.3% respectively. This reflects that traditional algorithms cannot fully optimize the scheduling of energy storage and power generation systems when faced with insufficient photovoltaic power generation, resulting in a significant reduction in system efficiency.

[0069] Figure 6The execution times of energy scheduling using GBDT+MPC, dynamic programming, and greedy algorithms are shown for systems of different sizes. The GBDT+MPC algorithm demonstrates a significant advantage across all system sizes, consistently maintaining the lowest execution time. In small-scale systems, GBDT+MPC's execution time is approximately 250ms, while the dynamic programming and greedy algorithms take 380ms and 300ms respectively, demonstrating GBDT+MPC's ability to respond quickly and efficiently to smaller tasks. As system size increases, GBDT+MPC maintains good scalability, with an execution time of only 850ms in very large-scale systems. In contrast, the dynamic programming algorithm takes as long as 1050ms, and the greedy algorithm reaches 940ms. This shows that GBDT+MPC adapts better to changes in system size and exhibits superior computational efficiency in large and very large-scale systems. While the dynamic programming algorithm has a relatively short execution time in small-scale systems, its execution time increases significantly with increasing system size. Especially in large-scale and ultra-large-scale systems, the execution time of the dynamic programming algorithm is significantly longer than that of the other two algorithms, at 820ms and 1050ms respectively. This phenomenon can be attributed to the exponential increase in computational cost when dynamic programming faces complex tasks, leading to a decrease in execution efficiency. Therefore, although the dynamic programming algorithm can theoretically find the global optimum, its high computational complexity significantly limits its scalability in practical applications, especially when dealing with large-scale system scheduling tasks, where it faces a significant bottleneck. The greedy algorithm performs relatively stably across all system sizes, and its execution time in small- to medium-scale systems is close to that of the GBDT+MPC algorithm, at 500ms and 300ms respectively. However, as the system size increases further, the execution time of the greedy algorithm also increases significantly, reaching 940ms in ultra-large-scale systems. Although the greedy algorithm is more efficient than the dynamic programming algorithm, its local optima mean that its scheduling performance may not be as good as the GBDT+MPC algorithm, especially when facing complex scheduling tasks, where the greedy algorithm cannot guarantee the validity of the global optimum.

[0070] In summary, the GBDT+MPC algorithm outperforms dynamic programming and greedy algorithms across systems of varying sizes. It maintains low execution time while adapting to system scaling, exhibiting strong robustness and stability. This result further validates the effectiveness of combining machine learning with predictive control in energy scheduling. In contrast, dynamic programming, due to its higher computational complexity, is more suitable for small-scale system scheduling, while greedy algorithms offer high execution efficiency in small to medium-scale systems, but their scheduling performance may have limitations, especially in very large-scale systems.

Claims

1. A method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system, characterized in that: The energy dispatch process includes the following steps: S1: Construct a multi-source data acquisition platform to collect and preprocess data including photovoltaic power generation data, energy storage equipment data, load demand data, and meteorological data. S2: Based on the preprocessed data in S1, the gradient boosting decision tree algorithm, i.e., the GBDT algorithm, is used to construct a photovoltaic power generation prediction model and a load demand prediction model. S3: Based on the output of the prediction model, a model predictive control algorithm, namely the MPC algorithm, is used for predictive control. The goal is to minimize the comprehensive cost including operating costs and energy storage losses. The charging and discharging strategy of the energy storage device is optimized in a rolling manner to generate the optimal control sequence in the rolling time domain. S4: Based on real-time grid time-of-use pricing, construct a grid-energy storage collaborative optimization model with the goal of minimizing grid interaction costs, and formulate a power purchase and sale strategy with the grid; S5: Based on system operation feedback, adaptively update the parameters of the two prediction models and control strategies in step S2.

2. The method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system according to claim 1, characterized in that: In step S2, a photovoltaic power generation prediction model and a load demand prediction model are constructed. These photovoltaic power generation prediction models are all based on the GBDT algorithm and are collectively referred to as GBDT energy prediction models. The specific construction steps are as follows: (S2.1) Integrate historical photovoltaic power generation data, meteorological characteristics, and load time series data to construct a multi-dimensional input feature set; (S2.2) Generate training label data based on the input features and construct a nonlinear regression model; (S2.3) Determine the optimal parameter combination for the decision tree through cross-validation and hyperparameter optimization; (S2.4) An iterative correction mechanism is adopted to generate weak learners in each round and optimize and correct the prediction error of the previous model until the convergence condition is met.

3. The method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system according to claim 1, characterized in that: In step S3, the MPC algorithm is used for predictive control, specifically including: (S3.1) First, define the system's state variables and control variables. Let the system's state vector be... It includes at least the photovoltaic power generation capacity, the charge / discharge state of the energy storage device, and the load demand, with the control vector being... If the system includes at least the charging and discharging power of the energy storage device and the load distribution, then the dynamic model of the system can be expressed as: (Official 1) in, It is the system's state transition function. This represents random perturbations in the system, reflecting the uncertainty of photovoltaic power generation. (S3.2) Next, define the control objective and constraints. The objective function can be expressed as: (Official 2) in, To predict the length of the time domain, and These are the costs of electricity purchase and energy storage losses, respectively. and The first The power purchase capacity and energy storage capacity at any given time; (S3.3) A rolling optimization and online update mechanism is adopted, with the prediction time domain set to 24 hours and the control time domain set to 1 hour.

4. The method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system according to claim 1, characterized in that: Step S4, constructing the grid-energy storage collaborative optimization model includes the following steps: (1) Construct a time-of-use pricing response function to characterize the economics of the interaction between the system and the power grid. The response function is as follows: (Formula 3) in, The power grid interaction cost function at time t; This indicates the power purchased from the power grid. This indicates the power output sold to the grid. and These represent the unit price of electricity purchase and sales, or the electricity price weighting factor, respectively, for the corresponding time period; (2) Design the priority rules for energy storage charging and discharging: prioritize charging operations during off-peak hours and prioritize discharging plans during peak hours; through the time-of-use pricing mechanism, minimize the purchase of electricity from the grid during peak hours and sell electricity to the grid using the energy storage system during off-peak hours, thereby improving the grid interaction economy of the system; (3) Set grid power interaction constraints: To meet system operational safety and power purchase capacity constraints; (4) Construct an optimization model with the goal of minimizing grid interaction costs, and use dynamic programming algorithm to solve for the optimal power purchase and sale strategy within the daily cycle range: (Formula 4).

5. The method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system according to claim 2, characterized in that: When constructing the GBDT energy forecasting model, it is assumed that photovoltaic power generation and load demand are denoted as follows: and Then, the following formula can be used for prediction: (Official 5) (Official 6) in, and These are the weights of the decision tree. and For the output of the decision tree, and The number of decision trees.

6. The method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system according to claim 5, characterized in that: It also includes the calculation process of minimizing the prediction error of the GBDT energy prediction model, including finding a function By progressively optimizing the loss function, the prediction error is minimized. This involves finding the function... The method is as follows: (1) Initialize the model: (Official 5) in, Represents the loss function. Let γ represent the target variable, and let γ represent the initial values ​​of the model parameters. (3) For each iteration Calculate the residuals using the following formula: (Official 6) Where: m represents the current iteration round, increasing sequentially from 1 to M, and M is the total number of iteration rounds; r im Let F(x) represent the residual of the i-th data sample in the m-th round, which is used to measure the deviation between the current model prediction and the target value; F(x) is the output prediction value of the model in the m-th round, which is updated based on the model results of the previous round. Describing the target variable F m-1 (x) represents the model's predicted output function in the (m-1)th round; (4) Fit a new decision tree This allows it to minimize the sum of squares of the residuals: (Official 7) (5) Update the model: (Official 8) in, This represents the output of the prediction model after the m-th iteration; This represents the model output from the previous iteration; In the first The weak learner (regression tree) obtained during training is used to fit the residuals. ; Let be the learning rate, where 0 < 1. ≤1, used to control the contribution ratio of each tree to the final model; Through the above iterative update process, the energy prediction model based on GBDT can gradually optimize the loss function. Ultimately, a model capable of accurately predicting photovoltaic power generation energy was obtained. .

7. The method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system according to claim 3, characterized in that: Step S3 involves rolling optimization of the energy storage device's charging and discharging strategy, based on the MPC control algorithm. This includes formulating an energy storage discharge optimization strategy, assuming the energy storage device's charging power is... The discharge power is The optimization objective is then expressed as: (Official 9) in, and The cost functions for charging and discharging are respectively. To optimize the time period, and These represent the charging and discharging power of the energy storage system at time t, respectively.

8. The method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system according to claim 6, characterized in that: It also includes an evaluation of the performance of GBDT-based energy prediction models, the evaluation methods including cross-validation or hyperparameter optimization, wherein the cross-validation method includes the following steps: (1) Dataset definition: The dataset is a complete dataset composed of multi-source data such as photovoltaic power generation systems, energy storage devices, and load demand, denoted as , where x i Represents the input features, y i Represent the target variable; (2) Cross-validation method: Divide the dataset into k subsets. For each fold i=1,2,....,k, select the i-th subset as the validation set and the remaining k-1 subsets as the training set. Train the GBDT model on the training set each time and evaluate the model performance on the validation set. Record the model performance index. After cross-validation is completed, calculate the average value of the validation results of each fold as the overall evaluation standard of the model performance. (3) Performance evaluation: After cross-validation is completed, the average value of all K rounds of validation results is calculated as the overall performance index of the model; The hyperparameter optimization method is as follows: before model training, the optimal combination of hyperparameters is searched to improve the performance of the model. The hyperparameter optimization methods include grid search and random search; the optimized hyperparameters include the maximum depth of the decision tree, the learning rate, and the subsample ratio.

9. The method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system according to claim 7, characterized in that: The constraints include physical and operational constraints of the system, including that the charging and discharging power of the energy storage device should meet its rated power limit, the state of the energy storage device should be within its capacity range, and the load demand should be met. The above three constraints are expressed as follows: Formula 10: (Official 10) in, and The minimum and maximum charge and discharge power of the energy storage device, and The minimum and maximum energy states of an energy storage device. For the first The load power requirement at any given time.

10. The method for efficient energy dispatch and intelligent control of a photovoltaic energy storage system according to claim 8, characterized in that: It also includes the prediction of photovoltaic power generation. The GBDT energy prediction model uses historical data and environmental factors to predict future photovoltaic power generation. Let the predicted photovoltaic power generation be... ,but: (Official 11) Furthermore, intelligent scheduling is achieved by combining the MPC algorithm. The MPC algorithm, by integrating historical data and environmental factors, predicts grid demand and photovoltaic power generation over a future period, thereby optimizing the charging and discharging strategies of energy storage devices. Specifically, formulas 12 and 13 describe the optimization objectives for power balance and scheduling. Formula 14 dynamically models the state update of energy storage devices based on their characteristics, realizing dynamic energy scheduling based on the innovative points of this invention. Let the predicted grid demand be... The MPC optimization problem can then be described as follows: (Official 12) Of which, photovoltaic power generation is The charging power of the energy storage device is The discharge power is The demand of the power grid is The charging and discharging efficiencies of the energy storage devices are respectively and ; The power grid demand is The equilibrium equation can be expressed as: (Official 13) The state update equation for the energy storage device is: (Official 14) Among them, the status of energy storage devices is Its capacity is ; Constraints: (Official 15) (Official 16) (Official 17) Formulas 15 and 16 define the dynamic range of the charging and discharging power of energy storage devices, while Formula 17 further considers the state of the energy storage device. and capacity To achieve real-time control of energy storage capacity and ensure the safety and efficiency of system operation, these constraints are combined with the existing modeling framework and further enhanced by the addition of dynamic scheduling objectives and innovative state update equations.

Images