An artificial intelligence-based distributed photovoltaic storage and operation method

Abstract

The application relates to the technical field of energy storage, and discloses a distributed photovoltaic energy storage matching and operation method based on artificial intelligence, which comprises the following steps: acquiring initial multi-source data sets of a distributed photovoltaic system on the user side; based on the initial multi-source data sets, performing feature engineering and gradient boosting tree classification model processing to obtain a matching necessity probability value and to judge whether the user side needs matching; when the user side needs matching, performing optimal energy storage configuration on the distributed photovoltaic system by using a bidirectional long short-term memory network model and an improved genetic algorithm, and obtaining a next-day prediction data set and a target configuration parameter set; based on the target configuration parameter set and the next-day prediction data set, performing dynamic self-adaptive regulation and control method processing to obtain a target charging and discharging strategy of the distributed photovoltaic system; and using the target charging and discharging strategy to control the operation of the distributed photovoltaic system, so that the matching operation result is obtained, and the renewable energy consumption efficiency and comprehensive income of the distributed photovoltaic matching project are improved.

Description

Technical Field

[0001] This invention relates to the field of energy storage technology, specifically to a distributed photovoltaic power generation and storage and operation method based on artificial intelligence. Background Technology

[0002] As an important form of clean energy utilization, the efficient absorption and economical operation of distributed photovoltaic power relies on the scientific configuration and scheduling of energy storage systems. However, existing technologies face the following core problems:

[0003] (1) The method of allocating storage lacks scientific basis: existing methods are mostly based on static data (such as historical averages), which fail to delve into the load-photovoltaic coupling law hidden in massive historical data. As a result, the decision of "allocating or not allocating" lacks data support, and the situation of "over-capacity" or "insufficient capacity" occurs from time to time, making it impossible to accurately quantify the necessity of allocating storage.

[0004] (2) Poor adaptability of photovoltaic-storage system operation strategy: Existing energy storage systems mostly adopt charging and discharging strategies based on fixed thresholds (such as "valley charging and peak discharging"), which cannot adaptively respond to complex real-time situations, such as sudden changes in electricity prices, abnormal weather, and sudden increases in load. It is difficult to maximize revenue while ensuring operational stability, and a dynamic control mechanism that takes into account both economic efficiency and battery health has not been formed.

[0005] (3) Fragmentation of AI technology application: In existing technologies, artificial intelligence is mostly used only in a single link (such as load forecasting), and has not formed a full-process intelligent closed loop of "data collection-storage allocation decision-operation optimization-feedback iteration". It is impossible to continuously improve the system adaptability through multi-model collaborative optimization. Although simple mathematical programming models or classic machine learning models (such as regression and clustering) can process numerical data, they are difficult to understand and utilize valuable unstructured text information such as market data, weather warnings, and equipment manuals. The decision-making dimension is single, lacks foresight, and cannot provide comprehensive data support for energy storage configuration and operation.

[0006] Energy storage systems are an important means of addressing the intermittency and volatility of distributed photovoltaic power, and can improve the flexibility and reliability of power systems. However, the current energy storage configuration lacks scientific methods, often resulting in problems such as unreasonable capacity allocation and low operating efficiency, leading to high investment costs and poor economic benefits. Summary of the Invention

[0007] This invention provides an artificial intelligence-based method for distributed photovoltaic energy storage and operation, which addresses the current lack of scientific methods for energy storage configuration, resulting in problems such as unreasonable capacity configuration and low operating efficiency.

[0008] In a first aspect, the present invention provides a distributed photovoltaic power generation and storage method based on artificial intelligence, the method comprising:

[0009] The process involves: acquiring initial multi-source datasets of the distributed photovoltaic (PV) system on the user side; obtaining the necessity probability value of energy storage based on the initial multi-source dataset through feature engineering and gradient boosting tree classification model; determining whether energy storage is needed on the user side based on the necessity probability value; when energy storage is needed on the user side, optimizing the energy storage configuration of the distributed PV system using a bidirectional long short-term memory (LSTM) network model and an improved genetic algorithm, and obtaining the next-day prediction dataset and target configuration parameter set. The bidirectional LSTM network model is obtained by stacking a forward LSTM network model and a backward LSTM network model; obtaining the target charging and discharging strategy of the distributed PV system based on the target configuration parameter set and the next-day prediction dataset through a dynamic adaptive control method; and controlling the operation of the distributed PV system using the target charging and discharging strategy to obtain the energy storage operation results.

[0010] This invention provides an AI-based distributed photovoltaic (PV) energy storage and operation method. Through feature engineering and gradient boosting tree classification models, it can mine the load-PV coupling patterns in the data, enabling quantitative determination of the necessity of energy storage allocation. This overcomes the subjectivity of traditional experience-based judgments and addresses the lack of scientific basis in existing energy storage allocation decisions. Furthermore, by determining whether users need energy storage based on the probability value of its necessity, it avoids resource waste and revenue loss caused by blindly allocating or not allocating energy storage, solving the problem of "over-capacity" or "insufficient capacity" in capacity configuration. Furthermore, by utilizing a bidirectional long short-term memory network model and an improved genetic algorithm, it can accurately predict future key parameters and achieve optimal configuration for full lifecycle revenue, balancing economic efficiency and practicality, reducing energy storage investment costs and improving economic benefits. Furthermore, through a dynamic adaptive control method, the target charge-discharge strategy can adapt to complex operating conditions, balancing energy storage economics and battery health. Furthermore, by using the target charge-discharge strategy to control the operation of the distributed PV system, it achieves coordinated and efficient operation of distributed PV and energy storage, improving the renewable energy consumption efficiency and overall revenue of distributed PV energy storage projects. At the same time, it provides a physical basis for the construction of virtual power plants on the regional power supply side, which helps to promote the high proportion of regional renewable energy consumption.

[0011] In one optional implementation, based on the initial multi-source dataset, after feature engineering and gradient boosting tree classification model processing, the probability value of storage necessity is obtained, including:

[0012] Data cleaning is performed on the multi-source dataset to obtain the target multi-source dataset; features are extracted from the target multi-source dataset to obtain a multi-dimensional feature parameter set; the multi-dimensional feature parameters are input into the gradient boosting tree classification model for processing to obtain the probability value of storage necessity.

[0013] The distributed photovoltaic (PV) energy storage and operation method based on artificial intelligence provided by this invention cleanses multi-source datasets, removing outliers and repairing missing values ​​to ensure data quality. Furthermore, feature extraction uncovers the core value of the data, forming multi-dimensional features adapted to a gradient boosting tree classification model, which helps improve the accuracy of determining the necessity of energy storage allocation. Moreover, by inputting the multi-dimensional feature parameters into the gradient boosting tree classification model, the probability value of energy storage necessity is obtained, achieving precise quantification of energy storage demand. This overcomes the subjectivity of traditional experience-based judgments and addresses the lack of scientific basis in existing energy storage allocation decisions.

[0014] In one optional implementation, when user-side energy storage is required, a bidirectional long short-term memory network model and an improved genetic algorithm are used to optimize the energy storage configuration of the distributed photovoltaic system, resulting in a next-day prediction dataset and a target configuration parameter set, including:

[0015] When the user side needs to allocate storage, the target optimization function and the user side's multi-dimensional dataset and next-day dataset are obtained. The target optimization function is used to represent maximizing the net benefit over the entire life cycle. The multi-dimensional dataset and next-day dataset are respectively input into the bidirectional long short-term memory network model for processing to obtain the multi-dimensional prediction dataset and the next-day prediction dataset. Based on the multi-dimensional prediction dataset and the preset constraint set, the target optimization function is solved using an improved genetic algorithm to obtain the target configuration parameter set.

[0016] The distributed photovoltaic energy storage and operation method based on artificial intelligence provided by this invention clarifies the target orientation and data support for energy storage configuration by acquiring the objective optimization function and multi-dimensional dataset, thus ensuring the economic efficiency of the configuration. Furthermore, the bidirectional long short-term memory network model can accurately predict future key parameters, providing forward-looking data for optimal capacity solution and improving the rationality of configuration. Moreover, based on the predicted dataset and constraints, and using an improved genetic algorithm, the optimal configuration parameters can be obtained within the compliance and feasibility boundaries, balancing economy, safety, and practicality, reducing energy storage investment costs and improving economic benefits.

[0017] In one optional implementation, based on the target configuration parameter set and the next day's forecast dataset, a target charging and discharging strategy for the distributed photovoltaic system is obtained through dynamic adaptive control methods, including:

[0018] Based on the target configuration parameter set and the next day's prediction dataset, a basic charging and discharging strategy for the distributed photovoltaic system is generated through deep reinforcement learning. The deep reinforcement learning method is used to control the balance between the net revenue and battery loss cost of the distributed photovoltaic system in a single period. The first real-time running dataset of the distributed photovoltaic system based on the basic charging and discharging strategy is obtained. Using the first real-time running dataset and the next day's prediction dataset, a target correction strategy is determined. The basic charging and discharging strategy is corrected in real time using the target correction strategy to obtain the target charging and discharging strategy.

[0019] The distributed photovoltaic power generation and storage operation method based on artificial intelligence provided by this invention generates a basic charging and discharging strategy through deep reinforcement learning, enabling advance planning of charging and discharging for the next day. This reduces battery loss while ensuring revenue during a single period. Furthermore, by acquiring a first real-time operating dataset and combining it with the next day's predicted dataset to determine a target correction strategy, the method can capture dynamic changes in actual operating conditions, accurately identify deviations between predictions and actual operating conditions, and clarify the direction of strategy adjustment. Moreover, the target correction strategy is used to correct the basic charging and discharging strategy in real time, improving the strategy's adaptability to real-time operating conditions and ensuring the economy and stability of the operation process.

[0020] In one optional implementation, the operation of the distributed photovoltaic system is controlled using a target charging and discharging strategy to obtain the distribution and storage operation results, including:

[0021] When the operation of a distributed photovoltaic system is controlled based on a target charging and discharging strategy, an artificial intelligence model is used to obtain the battery status of the distributed photovoltaic system. When the battery status is healthy, the operation of the distributed photovoltaic system is controlled using the target charging and discharging strategy to obtain the distribution and storage operation results. When the battery status is unhealthy, the operating parameters of the distributed photovoltaic system are adjusted based on the battery status, and the distribution and storage operation results are obtained.

[0022] The distributed photovoltaic (PV) power generation and storage operation method based on artificial intelligence provided by this invention acquires battery status through an AI model, enabling real-time monitoring of battery health and providing a safety basis for operation control. Furthermore, when the battery is healthy, it operates according to the target strategy, ensuring optimal economic operation of the distributed PV system under safe conditions. Moreover, when the battery is unhealthy, adjusting operating parameters can prevent further battery damage, extend battery life, and thus balance operational safety and sustainability.

[0023] In an optional implementation, the method further includes: updating the gradient boosting tree classification model, the bidirectional long short-term memory network model, the improved genetic algorithm, and the dynamic adaptive regulation method based on the storage operation results.

[0024] The distributed photovoltaic power allocation and storage operation method based on artificial intelligence provided by this invention updates various models and methods through the results of power allocation and storage operation, forming a closed-loop optimization of the entire process. This continuously improves the accuracy of power allocation and storage decisions, the rationality of energy storage configuration, and the adaptability of operation strategies, thereby enhancing the comprehensive benefits and generalization ability of the system in the long term.

[0025] Secondly, the present invention provides a distributed photovoltaic power generation and storage and operation device based on artificial intelligence, the device comprising:

[0026] The system comprises the following modules: an acquisition module for acquiring the initial multi-source dataset and the next-day prediction dataset of the distributed photovoltaic (PV) system on the user side; a first processing module for obtaining the probability value of energy storage necessity based on the initial multi-source dataset through feature engineering and a gradient boosting tree classification model; a judgment module for determining whether energy storage is needed on the user side based on the probability value of energy storage necessity; a configuration module for optimizing the energy storage configuration of the distributed PV system when energy storage is needed on the user side, using a bidirectional long short-term memory (LSTM) network model and an improved genetic algorithm to obtain the target configuration parameter set; the bidirectional LTM network model is obtained by stacking a forward LTM network model and a backward LTM network model; a second processing module for obtaining the target charging and discharging strategy of the distributed PV system based on the target configuration parameter set and the next-day prediction dataset through a dynamic adaptive control method; and a control module for controlling the operation of the distributed PV system using the target charging and discharging strategy to obtain the energy storage operation results.

[0027] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the artificial intelligence-based distributed photovoltaic power generation and storage and operation method of the first aspect or any corresponding embodiment described above.

[0028] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the artificial intelligence-based distributed photovoltaic power generation and storage and operation method of the first aspect or any corresponding embodiment described above.

[0029] Fifthly, the present invention provides a computer program product, including computer instructions, which are used to cause a computer to execute the artificial intelligence-based distributed photovoltaic power generation and storage and operation method of the first aspect or any corresponding embodiment described above. Attached Figure Description

[0030] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of an application scenario according to an embodiment of the present invention;

[0032] Figure 2 This is a flowchart illustrating a distributed photovoltaic power generation, storage, and operation method based on artificial intelligence according to an embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram of intelligent planning and configuration in an AI-based distributed photovoltaic power generation and storage and operation method according to an embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of intelligent operation optimization in an AI-based distributed photovoltaic power generation and storage and operation method according to an embodiment of the present invention;

[0035] Figure 5 This is a structural block diagram of a distributed photovoltaic power generation, storage and operation device based on artificial intelligence according to an embodiment of the present invention;

[0036] Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0038] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0039] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0040] As an optional application scenario of this invention, the specific application environment architecture or specific hardware architecture on which the execution of the AI-based distributed photovoltaic power generation, storage, and operation method depends is described herein. For example... Figure 1 As shown, the architecture system may include at least one terminal device and at least one server. Figure 1 The system is illustrated in the example, which includes a computer 101, a mobile terminal 102, and a server 103, and the terminal devices such as the computer 101 and the mobile terminal 102 are connected to the server 103 through a network 110.

[0041] Specifically, the terminal device can be a smartphone, tablet, laptop, PDA, desktop computer, game console, smart TV, smart wearable device, in-vehicle terminal, VR (Virtual Reality) device, AR (Augmented Reality) device, etc. Server 103 can be a standalone physical server, a server cluster, a distributed system, or a cloud server providing cloud services. Network 110 can be a wired or wireless network, examples of which include, but are not limited to, the Internet, corporate intranet, local area network, wide area network, mobile communication network, and combinations thereof.

[0042] This invention provides an artificial intelligence-based method for distributed photovoltaic (PV) energy storage and operation. First, it quantifies the necessity of energy storage through feature engineering and a gradient boosting tree classification model. Then, using a bidirectional long short-term memory network model and an improved genetic algorithm, it accurately predicts key future parameters and achieves optimal configuration for full lifecycle benefits. Finally, a dynamic adaptive control method enables the target charging and discharging strategy to adapt to complex operating conditions, balancing energy storage economics and battery health. This achieves coordinated and efficient operation of distributed PV and energy storage, improving energy consumption efficiency and overall benefits.

[0043] According to an embodiment of the present invention, an embodiment of a distributed photovoltaic power generation and storage and operation method based on artificial intelligence is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0044] This embodiment provides an artificial intelligence-based distributed photovoltaic power generation and storage operation method, which can be used in the aforementioned mobile terminals, such as mobile phones and tablets. Figure 2 This is a flowchart of an artificial intelligence-based distributed photovoltaic power generation, storage, and operation method according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps:

[0045] Step S201: Obtain the initial multi-source dataset of the distributed photovoltaic system on the user side.

[0046] In one optional embodiment, a distributed photovoltaic system refers to a small photovoltaic power supply system deployed on the user side, consisting of photovoltaic modules, inverters, and other equipment, which can generate solar power and meet the user's own electricity needs (surplus electricity can be fed into the grid).

[0047] In one optional embodiment, the initial multi-source dataset represents the user-side full-dimensional historical and basic data supporting the allocation and storage decision collection, with a time granularity of 5 minutes for industrial and commercial use, and may include:

[0048] 1. Load data: historical active power, peak-valley period distribution, daily / monthly load fluctuation coefficient;

[0049] 2. Photovoltaic data: installed capacity, historical output curve, self-consumption of electricity, grid connection of electricity and its proportion, and curtailment records;

[0050] 3. Market data: Time-of-use electricity pricing (peak / spot / valley time periods and prices), photovoltaic feed-in tariff, subsidy policies, and capacity tariffs;

[0051] 4. Environmental data: local meteorological data such as light intensity, temperature, and precipitation (last 3 years), and seasonal variation patterns.

[0052] In one optional embodiment, load data such as historical active power, peak-valley time distribution, and daily / monthly load fluctuation coefficient can be obtained through user-side electricity meters and energy management systems.

[0053] Furthermore, photovoltaic data such as installed capacity, historical output curves, self-generated and self-consumed electricity, grid-connected electricity and its proportion, and curtailment records can be extracted from photovoltaic inverters and metering equipment.

[0054] Furthermore, data from documents released by the local power authority, such as time-of-use tariffs (peak / short / valley periods and corresponding tariffs), photovoltaic feed-in tariffs, market policies, and capacity tariffs, can be compiled.

[0055] Furthermore, it can connect to the meteorological department's database to obtain meteorological data such as local sunlight intensity, temperature, and precipitation over the past three years, as well as seasonal variation patterns.

[0056] Step S202: Based on the initial multi-source dataset, the necessary probability value of storage is obtained through feature engineering and gradient boosting tree classification model.

[0057] In one optional embodiment, feature engineering refers to the preprocessing of cleaning and extracting core features from the initial multi-source dataset, with the aim of improving data quality, mining key information, and forming a feature parameter set that fits the model.

[0058] In one optional embodiment, the Gradient Boosting Decision Tree (GBDT) classification model represents a machine learning model based on the iterative ensemble of multiple decision trees. It optimizes the loss function through gradient descent to achieve binary classification decision of allocating or not allocating storage, and outputs the quantitative result of the necessity of allocating storage.

[0059] In one optional embodiment, full-dimensional data (including post-storage revenue data) of 1000+ distributed photovoltaic users are obtained, and those with storage as needed and those without storage as positive and negative samples, respectively, to form a model training dataset.

[0060] Furthermore, multi-dimensional characteristic parameters such as photovoltaic-load matching degree, peak-valley price difference coefficient, and curtailment loss rate are selected as input variables for the model.

[0061] Furthermore, basic hyperparameters such as the number of decision trees, initial learning rate, and depth of a single tree are set to initialize the first decision tree (using the mean of the sample labels as the initial prediction value). Then, in each iteration, the residual between the current model's predicted value and the true label of the sample is calculated, and a new decision tree is trained with the residual as the target. The loss function is minimized using gradient descent, the weights of the new decision tree are determined and integrated into the ensemble model, and the iteration is repeated until the preset number of trees is reached or the loss function converges, resulting in a trained gradient boosting tree classification model.

[0062] Furthermore, during the training process, a grid search algorithm can be used to optimize the combination of hyperparameters such as learning rate, tree depth, and number of leaf nodes. The classification accuracy is used as the evaluation index to determine the optimal combination of hyperparameters and ensure that the model classification accuracy is ≥90%.

[0063] In one optional embodiment, the probability value of energy storage necessity represents a 0-100% quantitative value output by the GBDT model, which is used to characterize the possibility that the benefits of configuring energy storage on the user side can cover the costs.

[0064] In one optional embodiment, by feature engineering to mine the core value of the initial multi-source data and using the GBDT classification model to quantify the necessity of storage, an objective and accurate probability value of the necessity of storage can be obtained.

[0065] In one optional embodiment, when performing feature engineering on the initial multi-source dataset, the Local Outlier Factor (LOF) algorithm is first used for data cleaning, and principal component analysis (PCA) is used to extract key features. Then, the extracted features are input into the gradient boosting tree classification model for processing, thereby enabling binary classification quantification of the necessity of storage.

[0066] Step S203: Determine whether the user side needs storage allocation based on the probability value of storage allocation necessity.

[0067] In one optional embodiment, based on the probability value of the necessity of energy storage, by presetting an objective decision threshold, it is possible to further determine whether the distributed photovoltaic system on the user side needs to be equipped with energy storage, replacing the traditional experience-based judgment, avoiding the waste of resources and loss of income caused by blindly allocating or not allocating energy storage, and ensuring the scientific and economical nature of energy storage allocation decisions.

[0068] In one optional embodiment, if the probability value of the necessity of storage is ≥50%, it is determined that storage is needed; otherwise, it is determined that storage is not needed.

[0069] In step S204, when the user side needs to allocate energy storage, the optimal energy storage configuration of the distributed photovoltaic system is performed using a bidirectional long short-term memory network model and an improved genetic algorithm, and the next day's prediction dataset and target configuration parameter set are obtained.

[0070] In one optional embodiment, photovoltaic output is affected by factors such as weather and seasons, exhibiting both diurnal periodic local dependence and seasonal long-term dependence. Therefore, this embodiment utilizes a bidirectional long short-term memory network model (Bi-LSTM model) to simultaneously capture the forward and backward dependencies of the time series. Furthermore, given the uncertainties in photovoltaic and load forecasting, the bidirectional structure helps the model learn confidence information from a complete temporal context.

[0071] In one optional embodiment, the bidirectional long short-term memory network model is obtained by stacking a forward long short-term memory network model and a backward long short-term memory network model. Furthermore, for each time step, the forward and backward hidden states are combined to form the final hidden state for that time step.

[0072] Furthermore, compared with the ordinary one-way LSTM model, the two-way LSTM model can better capture two-way dependency information (such as the impact of nighttime load on daytime energy storage decisions), making distributed photovoltaic energy storage more accurate and reliable.

[0073] In one optional embodiment, the input of the bidirectional long short-term memory network model may include parameters such as historical photovoltaic output, irradiance, temperature, historical load demand, weekday / holiday, time-of-use electricity price, and real-time electricity price; the output may include photovoltaic output forecast data, load demand forecast data, and electricity price forecast data, while also providing forecast uncertainty and extreme event probabilities.

[0074] In one alternative embodiment, the improved genetic algorithm represents an optimization algorithm that introduces an adaptive crossover and mutation operator on the basis of the traditional genetic algorithm, which is used to solve the optimal energy storage configuration under constraints, thereby improving optimization efficiency and result accuracy.

[0075] In one optional embodiment, the next-day prediction dataset represents the prediction results of key operating parameters for the next day generated by a bidirectional long short-term memory network (Bi-LSTM) model. It may include time-period load demand, photovoltaic output (prediction error ≤10%), time-of-use electricity price and fluctuation trend, and is used to formulate the day-ahead operation plan for energy storage.

[0076] In an optional embodiment, the target configuration parameter set represents the core configuration parameters of energy storage obtained by model calculation with the goal of maximizing net benefit over the entire life cycle, and may include the optimal energy storage capacity (Eopt) and the maximum charge and discharge power (Pmax).

[0077] In one optional embodiment, for users determined to require energy storage, long-term key parameters can be accurately predicted using a bidirectional long short-term memory network model. Then, using an improved genetic algorithm, under constraints such as charging and discharging power, SOC, and investment payback period, the target configuration parameter set that maximizes net income over the entire life cycle is solved. At the same time, a next-day prediction dataset is generated to support the formulation of the next-day operation strategy, thereby achieving a balance between the economy and practicality of energy storage configuration.

[0078] Step S205: Based on the target configuration parameter set and the next day's prediction dataset, the target charging and discharging strategy of the distributed photovoltaic system is obtained through dynamic adaptive control method.

[0079] In one optional embodiment, the dynamic adaptive control method represents a two-stage control method for realizing day-ahead planning and real-time correction. It can formulate strategies in advance based on forecast data and adaptively respond to real-time operating condition changes, balancing energy storage economy and battery health.

[0080] In one optional embodiment, the target charge and discharge strategy represents a quantitative operation command generated through two-stage regulation, which may include charge and discharge power and SOC control range for each time period. This can adapt to the predicted operating conditions for the next day and cope with real-time fluctuations, ensuring the economical and safe operation of the energy storage system.

[0081] In one optional embodiment, based on the optimal configuration parameters of energy storage and the prediction data of key operating conditions for the next day, a target charging and discharging strategy adapted to complex operating conditions can be dynamically generated through a two-stage control logic of daily planning and real-time dynamic correction. This solves the problem of poor adaptability of traditional fixed threshold strategies and achieves the dual goals of maximizing energy storage benefits and extending battery life.

[0082] Step S206: Using the target charging and discharging strategy, control the operation of the distributed photovoltaic system to obtain the distribution and storage operation results.

[0083] In one optional embodiment, the distribution and storage operation results represent the full-dimensional quantitative results generated during the operation of the distributed photovoltaic system under the coordinated control of the target charging and discharging strategy and the battery health management strategy. These results may include energy consumption-related data, economic benefit data, equipment operating status data, etc.

[0084] In one optional embodiment, the target charging and discharging strategy is transformed into specific operation control instructions for the distributed photovoltaic system. At the same time, the operating status is dynamically adjusted in conjunction with the battery health management strategy, thereby ultimately achieving the coordinated and efficient operation of photovoltaic and energy storage, and outputting full-dimensional operating results covering energy consumption, economic benefits, and equipment status.

[0085] The AI-based distributed photovoltaic (PV) energy storage and operation method provided in this embodiment, through feature engineering and gradient boosting tree classification model processing, can mine the load-PV coupling patterns in the data, realize the quantitative determination of the necessity of energy storage, overcome the subjectivity of traditional experience-based judgment, and solve the deficiency of existing energy storage decision-making lacking scientific basis. Furthermore, it determines whether users need energy storage based on the probability value of energy storage necessity, avoiding resource waste and revenue loss caused by blindly allocating or not allocating energy storage, and solving the problem of "over-capacity" or "insufficient capacity" in capacity configuration. Furthermore, by utilizing a bidirectional long short-term memory network model and an improved genetic algorithm, it can accurately predict future key parameters and achieve optimal configuration of full life cycle revenue, taking into account both economy and practicality, reducing energy storage investment costs and improving economic benefits. Furthermore, through a dynamic adaptive control method, the target charge and discharge strategy can adapt to complex operating conditions, balancing energy storage economics and battery health. Furthermore, by using the target charge and discharge strategy to control the operation of the distributed PV system, it achieves the coordinated and efficient operation of distributed PV and energy storage, improving the renewable energy consumption efficiency and overall revenue of distributed PV energy storage projects. At the same time, it provides a physical basis for the construction of virtual power plants on the regional power supply side, which helps to promote the high proportion of regional renewable energy consumption.

[0086] In some optional implementations, step S202 above includes:

[0087] Step S2021: Perform data cleaning on the multi-source dataset to obtain the target multi-source dataset.

[0088] In one optional embodiment, an outlier can be identified in load data, photovoltaic data, market data, and environmental data using an isolated forest algorithm. Then, the identified outliers are marked and removed to avoid interfering with data patterns.

[0089] Furthermore, missing values ​​in the multi-source dataset can be repaired using linear interpolation to ensure data continuity and integrity. Finally, after the above outlier removal and missing value repair, the corresponding target multi-source dataset can be obtained.

[0090] In an optional embodiment, outliers can also be identified using the DBSCAN clustering algorithm; missing values ​​can be repaired using the K-nearest neighbor (KNN) interpolation method to ensure the temporal correlation and accuracy of the data.

[0091] In an optional embodiment, a sliding window method (with a window size of 1 hour) can also be used to detect data continuity. Abnormal data with fluctuations exceeding 3 times the standard deviation of the same period in history within the window are marked, and invalid data is filtered in combination with industry standard thresholds. Missing values ​​are repaired using the seasonal trend decomposition (STL) interpolation method to adapt to environmental data and load data with obvious seasonal fluctuation characteristics.

[0092] Step S2022: Extract features from the target multi-source dataset to obtain a multi-dimensional feature parameter set.

[0093] In one optional embodiment, the multidimensional feature parameter set represents a set of features that accurately reflect the photovoltaic-load characteristics, economic benefit potential, and storage demand on the user side. It may include multidimensional feature parameters such as photovoltaic-load matching degree (the proportion of overlap between photovoltaic output and load), peak-valley price difference coefficient (peak-valley electricity price difference / flat-rate electricity price), and curtailment loss rate (curtailed photovoltaic power × grid connection price / total power generation × electricity sales price).

[0094] In one optional embodiment, by mining key information from target multi-source data, core features that can characterize the necessity of storage are constructed and a corresponding multi-dimensional feature parameter set is formed.

[0095] In an optional embodiment, the constructed features can also be standardized (uniformed in scale) to ensure that the weights of each feature are balanced during model training. Furthermore, redundant features can be filtered out.

[0096] Step S2023: Input the multidimensional feature parameters into the gradient boosting tree classification model for processing to obtain the probability value of storage necessity.

[0097] In one optional embodiment, the gradient boosting tree classification model, after training, performs calculations on the input multidimensional feature parameters and outputs a quantitative probability value that objectively reflects the possibility that the user's storage benefits cover the costs, i.e., the storage necessity probability value.

[0098] In some optional implementations, step S204 above includes:

[0099] Step S2041: When the user side needs to allocate storage, obtain the target optimization function and the user side's multi-dimensional dataset and the next day's dataset.

[0100] In an alternative embodiment, the objective optimization function is used to characterize maximizing the net benefit over the entire lifecycle, as shown in the following equation (1):

[0101] (1)

[0102] In the formula: express The added value of self-consumption during a certain period is (electricity sales price - grid connection price) × the amount of electricity stored to replace grid connection; express Peak-valley arbitrage profit during a given period (discharge price × discharge volume - charging price × charging volume). This represents the initial investment in energy storage (cost per unit capacity × capacity, approximately 0.8 yuan / Wh for lithium batteries). This represents the operation and maintenance cost (initial investment × 2% / year × lifespan). Indicates the life cycle (in 20 years).

[0103] In one optional embodiment, the multi-dimensional dataset represents the user-side full-dimensional basic data collected to support the calculation of optimal energy storage capacity (long-term configuration decision), which may include parameter information such as historical photovoltaic output, irradiance, temperature, historical load demand, weekdays / holidays, time-of-use electricity price, and real-time electricity price.

[0104] In one optional embodiment, the next day dataset refers to the basic data collected to support the formulation of the next day's operation strategy (short-term operation decision), which may include parameters such as photovoltaic output, irradiance, temperature, load demand, and electricity price for the next day.

[0105] Step S2042: Input the multi-dimensional dataset and the next day's dataset into the bidirectional long short-term memory network model for processing to obtain the multi-dimensional prediction dataset and the next day's prediction dataset.

[0106] In one optional embodiment, the multi-dimensional prediction dataset may include time-period load (error ≤ 10%), photovoltaic output (error ≤ 10%), and electricity price policy adjustment trends (such as peak-valley period variation probability, user electricity price increase, and on-grid electricity price in the electricity market) for the next year.

[0107] In one optional embodiment, the next day's forecast dataset may include the next day's 24-hour period load (error ≤ 10%), photovoltaic output (error ≤ 10%), and electricity price policy adjustment trends (such as peak-valley period variation probability, user electricity price increase, and electricity market on-grid price).

[0108] In one optional embodiment, a bidirectional long short-term memory network model is used to capture long-term data patterns and short-term operational characteristics of the next day, thereby generating corresponding accurate prediction data, namely a multi-dimensional prediction dataset and a next-day prediction dataset.

[0109] For example, after inputting the multi-dimensional dataset and the next day's dataset into the bidirectional long short-term memory network model, for the multi-dimensional dataset, the bidirectional long short-term memory network model predicts key parameters for each time period of the next year; for the next day's predicted dataset, the bidirectional long short-term memory network model predicts key parameters for each time period of the next 24 hours.

[0110] Step S2043: Based on the multi-dimensional prediction dataset and the preset constraint set, the improved genetic algorithm is used to solve the objective optimization function to obtain the objective configuration parameter set.

[0111] In one optional embodiment, a preset set of constraints is used to ensure the safety, feasibility, and economy of the energy storage configuration, and may include:

[0112] 1. The charging and discharging power constraints are shown in the following equations (2) and (3):

[0113] (2)

[0114] (3)

[0115] In the formula: Indicates the energy storage charging power; This refers to the excess power of a distributed photovoltaic (PV) system, which is the portion of PV output that exceeds the current load demand of users. Indicates the energy storage discharge power; This indicates the load deficit on the user side, which is the portion of the current load demand that exceeds the real-time output of the photovoltaic system.

[0116] 2. SOC constraint, as shown in the following relation (4):

[0117] (4)

[0118] In the formula: It indicates the state of charge of the energy storage battery, which reflects the proportion of the current remaining charge of the energy storage battery to its rated capacity.

[0119] 3. Economic constraints: Investment payback period ≤ 4 years.

[0120] In one optional embodiment, within a preset constraint boundary, an improved genetic algorithm is used to find energy storage configuration parameters that maximize the net benefit over the entire life cycle, thereby achieving a balance between the economy and feasibility of the configuration.

[0121] For example, firstly, a set of initial solutions for energy storage capacity (E) and charge / discharge power (P) is generated as the initial population of the algorithm. Then, each initial solution is substituted into the objective optimization function, and combined with preset constraints, the fitness of each solution, i.e., the net benefit over the entire life cycle, is calculated.

[0122] Furthermore, an adaptive crossover and mutation operator is introduced to generate a new generation of population through selection, crossover, and mutation operations. This process is repeated iteratively until the fitness converges or the preset number of iterations is reached. Then, the solution with the highest fitness, i.e., the target configuration parameter set, is selected and output.

[0123] In some optional implementations, step S205 above includes:

[0124] Step S2051: Based on the target configuration parameter set and the next day's prediction dataset, the basic charging and discharging strategy of the distributed photovoltaic system is generated through deep reinforcement learning.

[0125] In one alternative embodiment, a deep reinforcement learning method is used to control the balance between the net revenue per time period and the battery depletion cost of the distributed photovoltaic system.

[0126] In one optional embodiment, by using target configuration parameters as constraints and next-day prediction data as a basis, deep reinforcement learning is used to balance the net gain in a single period with battery loss, thereby generating a basic charging and discharging strategy that adapts to the predicted operating conditions of the next day.

[0127] In one optional embodiment, the state space is the energy storage SOC predicted for the next day, the photovoltaic output for each time period, the load demand, and the time-of-use electricity price; the action space is the charging and discharging power (-Pmax~Pmax, negative values ​​are for discharging); and the reward function is the net income per time period minus the battery loss cost (positively correlated with the depth and number of charging and discharging cycles).

[0128] Furthermore, a deep reinforcement learning (DQN) model was trained using historical data from over 1000 users as training samples. Then, the target configuration parameters (Eopt, Pmax) and the next day's prediction dataset were input, and the model iteratively learned the optimal action, thereby outputting a basic charging and discharging strategy.

[0129] Furthermore, the basic charging and discharging strategy determines the charging and discharging power and SOC control range for each time period of the next day. Its core logic is: prioritize charging during periods when photovoltaic output is greater than load and electricity prices are low; prioritize discharging during periods when load is greater than photovoltaic output and electricity prices are high.

[0130] Step S2052: Obtain the first real-time running dataset of the distributed photovoltaic system based on the basic charging and discharging strategy.

[0131] In one optional embodiment, the first real-time running dataset may include parameter information such as actual time-period photovoltaic output, user load demand, real-time electricity price, current SOC of energy storage, battery temperature, and actual charging and discharging power.

[0132] In one optional embodiment, a real-time data acquisition terminal is deployed, and when the distributed photovoltaic system is running according to the basic charging and discharging strategy, the operating data of the distributed photovoltaic system is collected in 5-minute time granularity to form a corresponding first real-time operating dataset.

[0133] Step S2053: Using the first real-time running dataset and the next day's forecast dataset, determine the target correction strategy.

[0134] In one alternative embodiment, the first real-time running data is compared with the next day's predicted data on a time-by-time basis, and deviation scenarios are identified.

[0135] Examples include actual photovoltaic output being more than 10% lower than predicted, real-time electricity prices exceeding predicted values ​​by 5%, and sudden load drops.

[0136] Furthermore, a target correction strategy can be determined based on a preset deviation-correction mapping rule.

[0137] For example, if the actual photovoltaic output is more than 10% lower than the forecast: reduce the charging amount and reserve capacity to cope with the nighttime load; if the real-time electricity price suddenly increases (5% higher than the forecast): discharge in advance to lock in high returns; if a sudden drop in load is detected (such as users shutting down production during holidays): suspend discharge and switch to off-peak charging.

[0138] Step S2054: The basic charging and discharging strategy is modified in real time using the target correction strategy to obtain the target charging and discharging strategy.

[0139] In one optional embodiment, the adjustment instructions in the target correction strategy are used to replace or modify the corresponding content of the basic charge-discharge strategy on a time-period basis. Then, it is verified whether the corrected strategy meets the target configuration parameter constraints and battery safety requirements.

[0140] Furthermore, the fused strategy is transformed into specific, executable quantitative control instructions and the corresponding modified target charge and discharge strategy is output.

[0141] In some optional implementations, step S206 above includes:

[0142] Step S2061: When the distributed photovoltaic system is controlled to operate based on the target charging and discharging strategy, the battery status of the distributed photovoltaic system is obtained using an artificial intelligence model.

[0143] In one optional embodiment, a lightweight neural network model is deployed at the edge and battery operation data is collected in 5-minute time granularity, which may include real-time operation data such as battery SOC, charge / discharge current, voltage, battery temperature, cycle count, and charge / discharge depth.

[0144] Furthermore, the model learns from the battery's health data throughout its entire lifecycle and outputs the battery's health status, which can be categorized into three types: healthy, sub-healthy, and unhealthy. It also provides a quantitative value for SOH (Battery Health).

[0145] The training samples can include the operating characteristics of new batteries, aged batteries, and faulty batteries.

[0146] Step S2062: When the battery is in a healthy state, the operation of the distributed photovoltaic system is controlled using the target charge and discharge strategy to obtain the distribution and storage operation results.

[0147] In an optional embodiment, when the battery is in a healthy state, the quantitative instructions (charge and discharge power at each time period, SOC control range) in the target charge and discharge strategy are sent to the execution device such as the energy storage converter (PCS) to control the energy storage system to charge and discharge according to the plan.

[0148] Furthermore, the real-time monitoring system ensures precise matching between photovoltaic output, load demand, and energy storage charging and discharging operations, without power conflicts or exceeding limits.

[0149] Furthermore, full-dimensional operational data can be collected at a 5-minute time granularity, including actual photovoltaic output, actual user load, energy storage charging and discharging capacity, SOC change curve, time-period revenue, and battery operating parameters, and corresponding energy storage operation results can be generated.

[0150] Step S2063: When the battery is in an unhealthy state, adjust the operating parameters of the distributed photovoltaic system based on the battery status and obtain the distribution and storage operation results.

[0151] In one optional embodiment, when the battery is in an unhealthy state, the corresponding adjustment rules can be determined based on the reasons for the unhealthiness output by the artificial intelligence model.

[0152] For example, adjustment rules may include:

[0153] 1. SOC control: The range is dynamically adjusted according to the number of battery cycles (20%-90% for new batteries, 30%-80% for aged batteries).

[0154] 2. Temperature protection: When the battery temperature is ≥35℃, reduce the charging and discharging power to 50%; when the temperature is ≥40℃, stop charging and discharging.

[0155] 3. Optimization of charge and discharge depth: Based on battery health (SOH), the depth of charge and discharge is limited when SOH≤80% (single charge and discharge depth≤50%).

[0156] Furthermore, operating parameters such as charging and discharging power and SOC range are adjusted according to matching rules, and temporary control commands are generated and issued for execution to ensure that the adjusted parameters meet the battery's safe operating boundaries. Furthermore, under the adjusted parameters, the photovoltaic and energy storage systems maintain basic coordinated operation, with a focus on monitoring changes in battery status to prevent abnormal escalation.

[0157] Furthermore, the system's operational data after adjustment is collected and corresponding allocation and storage operation results are generated.

[0158] In some optional implementations, the above method further includes: updating the gradient boosting tree classification model, the bidirectional long short-term memory network model, the improved genetic algorithm, and the dynamic adaptive regulation method based on the storage operation results.

[0159] In one optional embodiment, by inputting the actual storage allocation and operation results back into each core model and control method, and dynamically updating the model parameters and optimizing the algorithm logic, the accuracy of storage allocation decisions, the economy of capacity configuration, and the adaptability of operation strategies can be continuously improved, thereby achieving continuous iterative upgrades of system performance.

[0160] In one optional embodiment, data such as actual storage and distribution revenue, curtailment rate changes, and actual load-photovoltaic matching degree are extracted from the results of this storage and distribution operation. Combined with the multidimensional feature parameters of the original training samples, new labeled samples are added to expand the training dataset.

[0161] Furthermore, the expanded dataset is input into the original GBDT model, and incremental training is used to adjust the weights and split nodes of the decision tree based on the new samples. Additionally, hyperparameters such as the learning rate and tree depth are re-optimized through grid search to minimize the classification loss function for the new sample set.

[0162] Furthermore, the classification accuracy of the updated model is verified using a test set. If the accuracy is met, the original model is replaced, and the threshold calibration logic for the necessity decision of storage is updated simultaneously.

[0163] In one optional embodiment, the actual data in the storage operation results are compared with the original predicted data of the model to calculate the prediction error for each time period, and then the abnormal time period data with an error exceeding 10% are screened to analyze the cause of the error.

[0164] Furthermore, the filtered error data, actual operating data, and newly added historical data (such as operating records under extreme weather conditions) are integrated into the training set to fine-tune the Bi-LSTM model and optimize hyperparameters such as the number of hidden network units and the number of iterations, thereby improving the model's ability to capture extreme operating conditions and policy changes.

[0165] Furthermore, the updated model is used to re-predict future short-term data to verify whether the prediction error is ≤10%; if it meets the standard, the updated model parameters are fixed and used for subsequent multi-dimensional prediction and next-day prediction dataset generation.

[0166] In one optional embodiment, the parameter weights of the objective optimization function are adjusted based on the actual net income and investment payback period data from the storage operation results. Furthermore, if constraint boundary adaptation problems occur during actual operation, the boundary values ​​of the preset constraint set are fine-tuned.

[0167] Furthermore, the deviation between the original solution (target configuration parameter set) of the improved genetic algorithm and the actual optimal parameters is statistically analyzed, and the probability threshold of the adaptive crossover and mutation operator is adjusted. Additionally, if low optimization efficiency exists, the diversity of population initialization is increased to shorten the convergence time.

[0168] Furthermore, the updated algorithm is used to resolve the objective optimization function, and the deviation between the simulated and actual returns corresponding to the output objective configuration parameter set is verified to be ≤5%. Once the target is met, the updated algorithm is officially implemented.

[0169] In one optional embodiment, updating the dynamic adaptive control method may include:

[0170] (1) The plan is to optimize the rule update: analyze the adaptation difference between the basic charging and discharging strategy and the actual working conditions in the storage operation results, adjust the weight of the reward function of deep reinforcement learning, and expand the subdivision of the action space.

[0171] (2) Real-time correction rule iteration: Statistically analyze the actual correction effect of different deviation scenarios (such as insufficient photovoltaic output and sudden load drop), add correction rules corresponding to the high-frequency deviation types, and optimize the input features of the lightweight neural network to improve the response speed of deviation identification and correction.

[0172] (3) Battery health management strategy optimization: Based on the actual health status changes of the battery in the storage and distribution operation results, adjust the health management parameters and update the dynamic adjustment logic of the SOC control range.

[0173] (4) Update and monitoring: Integrate the updated control rules and strategies into the two-stage operation framework, monitor the operation effect through short-term trial operation, and update the dynamic adaptive control method after confirming that the target is met.

[0174] In one example, an AI-based distributed photovoltaic energy storage and operation method is provided. This method aims to scientifically solve the "necessity" and "economic efficiency" issues of energy storage configuration through data-driven intelligent algorithms. Based on this, advanced AI models are used to achieve adaptive, forward-looking, and economical operation of the energy storage system, ultimately achieving the goal of the lowest cost and the highest benefit over the entire life cycle.

[0175] Specifically, this example addresses the problems of existing distributed photovoltaic (PV) energy storage methods lacking scientific basis, poor adaptability of PV-energy storage system operation strategies, and fragmented application of AI technology, by using artificial intelligence technology to achieve the following:

[0176] (1) Based on historical data such as user load, photovoltaic characteristics, and electricity price policies, and investment economic and technical analysis models, a machine learning classification model is used to accurately determine whether energy storage needs to be configured and its capacity.

[0177] (2) By combining real-time data and artificial intelligence algorithms, the optimal economic operation strategy is dynamically generated to maximize energy storage benefits and extend battery life.

[0178] In an alternative embodiment, such as Figure 3 and Figure 4 As shown in the example, this example provides an AI-based distributed photovoltaic power generation and storage (PV power generation and storage) operation method that constructs a "data-driven - intelligent decision-making - dynamic optimization" technical framework, including an intelligent power generation and storage decision-making module and an adaptive operation optimization module. The specific steps are as follows:

[0179] 1. Intelligent storage allocation decision module (solves the questions of "whether to allocate storage" and "how much capacity to allocate").

[0180] 1.1 Multi-source data acquisition and feature engineering: Acquiring key user-side data (time granularity: 5 minutes for industrial and commercial applications):

[0181] 1.1.1 Load data: historical active power, peak-valley period distribution, daily / monthly load fluctuation coefficient;

[0182] 1.1.2 Photovoltaic data: installed capacity, historical output curve, self-consumption of electricity, grid connection of electricity and its proportion, and curtailment records;

[0183] 1.1.3 Market Data: Time-of-use electricity pricing (peak / short / valley periods and prices), photovoltaic feed-in tariff, subsidy policies, and capacity tariffs;

[0184] 1.1.4 Environmental data: local meteorological data such as light intensity, temperature, and precipitation (last 3 years), and seasonal variation patterns.

[0185] 1.2 Feature Engineering Processing:

[0186] 1.2.1 Data cleaning: The Isolation Forest algorithm is used to identify outliers, and missing values ​​are repaired by linear interpolation;

[0187] 1.2.2 Feature Extraction: Construct a core feature set, including multi-dimensional feature parameters such as "PV-load matching degree" (the proportion of overlap between PV output and load), "peak-valley price difference coefficient" (peak-valley electricity price difference / flat-rate electricity price), and "curtailment loss rate" (curtailed PV power × grid connection price / total power generation × electricity sales price).

[0188] 1.3 Machine learning-based determination of storage necessity.

[0189] Construct a gradient boosting tree (GBDT) classification model to achieve a binary classification decision of "allocate storage / not allocate storage":

[0190] (1) Input: Multidimensional feature parameters extracted in Section 1.2.2;

[0191] (2) Output: Probability of necessity for allocation of reserves (0-100%);

[0192] (3) Decision threshold: if the probability is ≥50%, it is determined as "reservoir allocation is required" (the revenue from reserve allocation covers the cost); otherwise, it is "reservoir allocation is not required".

[0193] Model training and optimization:

[0194] (1) Training samples: 1000+ distributed photovoltaic user data (including data on revenue after storage), including positive samples (requiring storage) and negative samples (not requiring storage).

[0195] (2) Model optimization: The hyperparameters (learning rate, tree depth, etc.) are tuned by grid search to achieve a classification accuracy of ≥90%.

[0196] 1.4 Optimal Capacity Calculation Based on Deep Learning. For users determined to "require energy storage," the optimal energy storage capacity is calculated using the following steps:

[0197] 1.4.1 Multi-dimensional Prediction: In distributed photovoltaic power generation and storage systems, a bidirectional LSTM model is used to predict key parameters for the next year:

[0198] (1) Time-period load (error ≤ 10%), photovoltaic output (error ≤ 10%);

[0199] (2) Electricity price policy adjustment trends (such as the probability of changes during peak and off-peak periods, the increase in user electricity prices, and the on-grid electricity price in the electricity market).

[0200] The bidirectional LSTM model can be viewed as a stack of two unidirectional LSTMs, one forward and one backward. At each time step, the forward and backward hidden states are combined to form the final hidden state for that time step. The reasons for using this model are mainly twofold;

[0201] (1) Photovoltaic power output is affected by factors such as weather and seasons, and has both daily periodic local dependence and seasonal long-term dependence. Bidirectional LSTM can capture forward and backward dependence at the same time.

[0202] (2) There are uncertainties in photovoltaic and load forecasting. The two-way structure helps the model learn confidence information from the complete time context.

[0203] The model's inputs include parameters such as historical photovoltaic output, solar irradiance, temperature, historical load demand, weekday / holiday, time-of-use pricing, and real-time pricing. The model's outputs include photovoltaic output forecast data, load demand forecast data, and electricity price forecast data, while also providing forecast uncertainty and the probability of extreme events.

[0204] Compared with ordinary one-way LSTM models, two-way LSTM models can better capture two-way dependency information (such as the impact of nighttime load on daytime energy storage decisions), making distributed photovoltaic energy storage more accurate and reliable.

[0205] 1.4.2 Capacity optimization model: With the goal of “maximizing net income over the entire life cycle”, an optimization function is constructed, as shown in the above relation (1).

[0206] 1.4.3 Capacity Calculation:

[0207] An improved genetic algorithm (introducing an adaptive crossover mutation operator) is used to solve for the optimal capacity Eopt, where the constraints include charging and discharging power constraints (the above relations (2) and (3)), SOC constraints (the above relations (4)) and economic constraints (investment payback period ≤ 4 years).

[0208] 2. Adaptive operation optimization module (based on configuration dynamic operation strategy).

[0209] 2.1 Two-stage operation strategy framework. Combining reinforcement learning and real-time data, dynamic control is achieved through "daily plan + real-time correction":

[0210] Optimization plan prior to 2.1.1:

[0211] Based on the next-day forecast data (load, photovoltaic output, electricity price) using the Bi-LSTM model, a basic strategy is generated using deep reinforcement learning (DQN):

[0212] (1) State space: predicted SOC, photovoltaic output, load, and electricity price;

[0213] (2) Operation space: charging and discharging power (-Pmax~Pmax, negative value is discharge);

[0214] (3) Reward function: Net income per period - battery loss cost (positively correlated with charge / discharge depth and number of times).

[0215] (4) Core logic: Prioritize charging during periods when "PV output > load" and electricity prices are low; prioritize discharging during periods when "load > PV output" and electricity prices are high.

[0216] 2.1.2 Real-time Dynamic Correction: A lightweight neural network is deployed at the edge to collect real-time data (actual load, photovoltaic output, sudden weather changes) every 5 minutes to correct the charging and discharging plan.

[0217] If the actual photovoltaic output is more than 10% lower than the forecast: reduce the charging amount and reserve capacity to cope with the nighttime load; if the real-time electricity price suddenly increases (5% higher than the forecast): discharge in advance to lock in high returns; if a sudden drop in load is detected (such as users shutting down production during holidays): suspend discharge and switch to off-peak charging.

[0218] 2.2 Battery health management strategy.

[0219] Real-time battery status monitoring via AI models and dynamic adjustment of operating parameters: Based on the operational optimization in section 2.1, real-time battery monitoring and management are performed. The battery will exhibit certain boundary operating conditions, such as frequent charging and discharging leading to overheating and unsafe battery conditions. In practice, this involves optimizing the operational status while also considering the battery's tolerance. Specifically, this includes:

[0220] 1. SOC control: The range is dynamically adjusted according to the number of battery cycles (20%-90% for new batteries, 30%-80% for aged batteries).

[0221] 2. Temperature protection: When the battery temperature is ≥35℃, reduce the charging and discharging power to 50%; when the temperature is ≥40℃, stop charging and discharging.

[0222] 3. Optimization of charge and discharge depth: Based on battery health (SOH), the depth of charge and discharge is limited when SOH≤80% (single charge and discharge depth≤50%).

[0223] The AI-based distributed photovoltaic power generation and storage method for operation provided in this example has the following advantages:

[0224] (1) Precise decision-making on energy storage allocation: Based on historical data such as user load, photovoltaic characteristics, and electricity price policies, and investment economic and technical analysis models, the system uses machine learning classification models to accurately determine whether energy storage needs to be configured and its capacity.

[0225] (2) Maximizing economic benefits: The adaptive operation strategy increases the net return of energy storage over the entire life cycle by more than 20% compared with the traditional method, and shortens the investment payback period by more than 2 years;

[0226] (3) Strong adaptability and generalization: The AI ​​model can learn the electricity price policies and user load characteristics of different regions on its own, and adapt to various scenarios of different types of users;

[0227] (4) Extended battery life: Through health management strategies, the battery cycle life is increased by more than 20%.

[0228] This embodiment also provides an artificial intelligence-based distributed photovoltaic power generation, storage, and operation device, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0229] This embodiment provides a distributed photovoltaic power generation, storage, and operation device based on artificial intelligence, such as... Figure 5 As shown, the device includes:

[0230] The acquisition module 501 is used to acquire the initial multi-source dataset and the next day's prediction dataset of the distributed photovoltaic system on the user side.

[0231] The first processing module 502 is used to obtain the probability value of storage necessity based on the initial multi-source dataset through feature engineering and gradient boosting tree classification model.

[0232] The judgment module 503 is used to determine whether the user side needs to allocate storage based on the probability value of the storage necessity.

[0233] The configuration module 504 is used to perform optimal energy storage configuration for the distributed photovoltaic system when the user side needs to configure energy storage, using a bidirectional long short-term memory network model and an improved genetic algorithm to obtain a target configuration parameter set. The bidirectional long short-term memory network model is obtained by stacking a forward long short-term memory network model and a reverse long short-term memory network model.

[0234] The second processing module 505 is used to obtain the target charging and discharging strategy of the distributed photovoltaic system by processing the target configuration parameter set and the next day's prediction dataset through a dynamic adaptive control method.

[0235] The control module 506 is used to control the operation of the distributed photovoltaic system using the target charging and discharging strategy to obtain the distribution and storage operation results.

[0236] The AI-based distributed photovoltaic power generation, storage, and operation device provided in this embodiment of the invention can execute the AI-based distributed photovoltaic power generation, storage, and operation method provided in any embodiment of the invention, and possesses the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the above modules and units are the same as in the corresponding embodiments described above, and will not be repeated here.

[0237] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0238] The following is a detailed reference. Figure 6 This diagram illustrates a suitable structural design for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0239] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0240] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in the AI-based distributed photovoltaic power generation and storage operation method of the embodiments of the present invention.

[0241] Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0242] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the artificial intelligence-based distributed photovoltaic power generation and storage operation method shown in the above embodiments is implemented.

[0243] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0244] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A distributed photovoltaic power generation and storage system operation method based on artificial intelligence, characterized in that, The method includes: Obtain the initial multi-source dataset of the distributed photovoltaic system on the user side; Based on the initial multi-source dataset, after feature engineering and gradient boosting tree classification model processing, the probability value of storage necessity is obtained. The feature engineering refers to the preprocessing process of data cleaning and core feature extraction of the initial multi-source dataset, which is used for photovoltaic-load matching degree, peak-valley price difference coefficient, and curtailment loss rate. The gradient boosting tree classification model is trained with labeled samples of distributed photovoltaic users and the hyperparameters are optimized through grid search. The probability value of storage necessity represents the 0-100% quantitative value output by the gradient boosting tree classification model, which is used to characterize the possibility that the benefits of configuring energy storage on the user side can cover the costs. Determine whether the user side needs storage based on the probability value of storage necessity. When the user side requires energy storage, a bidirectional long short-term memory network model and an improved genetic algorithm are used to optimize the energy storage configuration of the distributed photovoltaic system, resulting in a next-day prediction dataset and a target configuration parameter set, including: When the user side needs to allocate storage, the target optimization function and the user side's multi-dimensional dataset and next day's dataset are obtained. The target optimization function is used to characterize the maximum net income over the entire life cycle. The multi-dimensional dataset and the next-day dataset are respectively input into the bidirectional long short-term memory network model for processing to obtain the multi-dimensional prediction dataset and the next-day prediction dataset. The bidirectional long short-term memory network model is obtained by stacking a forward long short-term memory network model and a backward long short-term memory network model. The bidirectional long short-term memory network model can simultaneously capture the forward and backward dependencies of the time series and is used to output photovoltaic power output prediction data, load demand prediction data, electricity price prediction data, prediction uncertainty and extreme event probability. Based on the multi-dimensional prediction dataset and the preset constraint set, the improved genetic algorithm is used to solve the objective optimization function to obtain the objective configuration parameter set. The preset constraint set is used to ensure the safety, feasibility and economy of energy storage configuration, including charging and discharging power constraints, SOC constraints and economic constraints. The objective configuration parameter set includes the optimal energy storage capacity and the maximum charging and discharging power. Based on the target configuration parameter set and the next day's prediction dataset, the target charging and discharging strategy of the distributed photovoltaic system is obtained through dynamic adaptive control method. By utilizing the target charging and discharging strategy, the operation of the distributed photovoltaic system is controlled to obtain the distribution and storage operation results; The objective optimization function is expressed as the following relation: In the formula: express The added value of self-consumption during a certain period is (electricity sales price - grid connection price) × the amount of electricity stored to replace grid connection; express Peak-valley arbitrage profit during a given period is calculated as: discharge price × discharge volume - charging price × charging volume. Indicates the initial investment in energy storage; Indicates operation and maintenance costs; Indicates the lifecycle.

2. The method according to claim 1, characterized in that, Based on the initial multi-source dataset, after feature engineering and gradient boosting tree classification model processing, the probability value of storage necessity is obtained, including: The multi-source dataset is cleaned to obtain the target multi-source dataset; Feature extraction is performed on the target multi-source dataset to obtain a multi-dimensional feature parameter set; The multidimensional feature parameters are input into the gradient boosting tree classification model for processing to obtain the probability value of the necessity of storage.

3. The method according to claim 1, characterized in that, Based on the target configuration parameter set and the next-day prediction dataset, the target charging and discharging strategy of the distributed photovoltaic system is obtained through dynamic adaptive control methods, including: Based on the target configuration parameter set and the next day prediction dataset, the basic charging and discharging strategy of the distributed photovoltaic system is generated through deep reinforcement learning. The deep reinforcement learning method is used to control the balance between the single-period net income and battery loss cost of the distributed photovoltaic system. Obtain the first real-time running dataset of the distributed photovoltaic system based on the basic charging and discharging strategy; Using the first real-time running dataset and the next day's predicted dataset, a target correction strategy is determined; The target charging and discharging strategy is obtained by using the target correction strategy to correct the basic charging and discharging strategy in real time.

4. The method according to claim 1, characterized in that, Using the target charge and discharge strategy, the operation of the distributed photovoltaic system is controlled to obtain the distribution and storage operation results, including: When the distributed photovoltaic system is controlled to operate based on the target charging and discharging strategy, the battery status of the distributed photovoltaic system is obtained using an artificial intelligence model; When the battery is in a healthy state, the operation of the distributed photovoltaic system is controlled using the target charging and discharging strategy to obtain the distribution and storage operation results; When the battery is in an unhealthy state, the operating parameters of the distributed photovoltaic system are adjusted based on the battery state, and the distribution and storage operation results are obtained.

5. The method according to claim 4, characterized in that, The method further includes: Based on the results of the storage allocation operation, the gradient boosting tree classification model, the bidirectional long short-term memory network model, the improved genetic algorithm, and the dynamic adaptive regulation method are updated respectively.

6. A distributed photovoltaic power generation, storage, and operation device based on artificial intelligence, characterized in that, The device includes: The acquisition module is used to acquire the initial multi-source dataset and the next day's prediction dataset of the distributed photovoltaic system on the user side; The first processing module is used to obtain the probability value of energy storage necessity based on the initial multi-source dataset through feature engineering and gradient boosting tree classification model. The feature engineering refers to the preprocessing process of data cleaning and core feature extraction of the initial multi-source dataset, which is used for photovoltaic-load matching degree, peak-valley price difference coefficient, and curtailment loss rate. The gradient boosting tree classification model is trained with labeled samples of distributed photovoltaic users and the hyperparameters are optimized through grid search. The probability value of energy storage necessity represents the 0-100% quantitative value output by the gradient boosting tree classification model, which is used to characterize the possibility that the benefits of configuring energy storage on the user side can cover the costs. The judgment module is used to determine whether the user side needs to allocate storage based on the probability value of storage necessity. The configuration module is used to perform optimal energy storage configuration for the distributed photovoltaic system when the user side requires energy storage, using a bidirectional long short-term memory network model and an improved genetic algorithm to obtain a target configuration parameter set, including: When the user side needs to allocate storage, the target optimization function and the user side's multi-dimensional dataset and next day's dataset are obtained. The target optimization function is used to characterize the maximum net income over the entire life cycle. The multi-dimensional dataset and the next-day dataset are respectively input into the bidirectional long short-term memory network model for processing to obtain the multi-dimensional prediction dataset and the next-day prediction dataset. The bidirectional long short-term memory network model is obtained by stacking a forward long short-term memory network model and a backward long short-term memory network model. The bidirectional long short-term memory network model can simultaneously capture the forward and backward dependencies of the time series and is used to output photovoltaic power output prediction data, load demand prediction data, electricity price prediction data, prediction uncertainty and extreme event probability. Based on the multi-dimensional prediction dataset and the preset constraint set, the improved genetic algorithm is used to solve the objective optimization function to obtain the objective configuration parameter set. The preset constraint set is used to ensure the safety, feasibility and economy of energy storage configuration, including charging and discharging power constraints, SOC constraints and economic constraints. The objective configuration parameter set includes the optimal energy storage capacity and the maximum charging and discharging power. The second processing module is used to obtain the target charging and discharging strategy of the distributed photovoltaic system by processing the target configuration parameter set and the next day's prediction dataset through a dynamic adaptive control method. The control module is used to control the operation of the distributed photovoltaic system using the target charging and discharging strategy, and to obtain the distribution and storage operation results; The objective optimization function is expressed as the following relation: In the formula: express The added value of self-consumption during a certain period is (electricity sales price - grid connection price) × the amount of electricity stored to replace grid connection; express Peak-valley arbitrage profit during a given period is calculated as: discharge price × discharge volume - charging price × charging volume. Indicates the initial investment in energy storage; Indicates operation and maintenance costs; Indicates the lifecycle.

7. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the distributed photovoltaic power generation and storage and operation method based on artificial intelligence as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the artificial intelligence-based distributed photovoltaic power generation and storage and operation method as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, It includes computer instructions for causing a computer to execute any one of claims 1 to 5, which is based on artificial intelligence and includes distributed photovoltaic power generation and storage and operation methods.

Images