AI prediction-based light storage and charging dynamic power distribution control method and system

Abstract

The application discloses a kind of light storage fills dynamic power allocation control method and system based on AI prediction, the method is collected photovoltaic, energy storage, charging, power grid and environmental multidimensional operation data by sensor network, and historical and real-time data set is obtained after pre-processing;Utilize historical data set to train AI joint prediction model, output photovoltaic output and charging load demand prediction result;Input system operation constraint condition parameter and demarcate optimization boundary;In combination with prediction result and constraint condition, each power module power instruction reference value is solved by multi-objective optimization algorithm, realize global optimal distribution;Based on hardware operation parameter and prediction and real-time value deviation correction reference value, issue instruction and monitor operation, trigger fault protection.The application effectively solves the problems of traditional system power scheduling rigidity, low photovoltaic consumption rate, large energy storage life loss, significant power grid impact, insufficient prediction accuracy, high energy transmission loss and weak multi-objective coordination.

Description

Technical Field

[0001] This invention belongs to the field of new energy power system control and artificial intelligence application technology, specifically relating to a dynamic power allocation control method and system for photovoltaic-storage-charging based on AI prediction. Background Technology

[0002] Photovoltaic power generation, as an important form of renewable energy utilization, has been widely used due to its clean and environmentally friendly advantages. The rapid growth in demand for electric vehicle charging has driven the large-scale deployment of integrated photovoltaic-energy storage-charging systems. This system integrates photovoltaic power generation, energy storage, and electric vehicle charging functions, enabling on-site energy production and consumption, reducing energy transmission losses, and mitigating the concentrated impact of charging loads on the power grid. It has become an important development direction in the field of energy utilization.

[0003] However, current photovoltaic-storage-charging systems still face numerous technical challenges during operation, which restrict their operational efficiency and reliability. First, photovoltaic power generation is significantly affected by natural factors such as solar irradiance, ambient temperature, and cloud movement, resulting in strong intermittency, fluctuation, and randomness in output power. It may experience sharp drops in a short period of time. Meanwhile, electric vehicle charging loads are characterized by uneven temporal distribution and uncertain demand, leading to prominent spatiotemporal mismatch problems between the "source" and "load" sides. It is difficult to maintain real-time power balance of the system, photovoltaic curtailment occurs frequently, and energy utilization efficiency is low.

[0004] Secondly, traditional power distribution control often employs fixed thresholds or simple rule-based strategies, lacking the ability to predict photovoltaic output and charging load in advance. This makes it difficult to effectively coordinate the multi-objective requirements of photovoltaic absorption, energy storage lifespan protection, and operational economy. Such strategies are poorly adaptable to changes in operating conditions. When photovoltaic power fluctuates or charging load changes abruptly, it is difficult to quickly adjust the energy storage charging and discharging interaction strategy with the grid, potentially causing problems such as DC bus voltage fluctuations and grid frequency deviations, affecting the stable operation of the system.

[0005] Furthermore, existing prediction models often rely on single-dimensional data, failing to fully integrate multi-source heterogeneous data such as historical operational data, meteorological information, and user charging habits. This results in insufficient prediction accuracy, with short-term prediction errors frequently exceeding 15%. Simultaneously, the models lack adaptive update mechanisms, making them susceptible to continuously widening prediction deviations due to environmental changes and load characteristics over long-term operation, thus affecting the rationality of power allocation decisions. In addition, the charging and discharging control of energy storage systems lacks refined optimization of the state of charge; frequent deep charging and discharging accelerate battery lifespan degradation, while the constraint control of grid-connected power is inaccurate, potentially triggering reverse current risks and grid protection actions. Moreover, the various devices in photovoltaic-storage-charging systems often come from different manufacturers, with inconsistent communication protocols, leading to data fragmentation and a lack of a unified energy management core. Traditional control methods struggle to integrate multi-dimensional operational data from the photovoltaic, energy storage, charging, and grid sides, and optimization algorithms face a contradiction between decision accuracy and real-time response speed in engineering applications, failing to meet the real-time control requirements of dynamic power allocation. Summary of the Invention

[0006] To address these issues, this invention provides a dynamic power allocation control method and system for photovoltaic-storage-charging systems based on AI prediction, which solves the problems of rigid power scheduling, low photovoltaic absorption rate, large energy storage life loss, significant grid impact, insufficient prediction accuracy, high energy transmission loss, and weak multi-objective coordination in traditional photovoltaic-storage-charging systems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction, comprising the following steps:

[0008] (1) Data acquisition and preprocessing: Multi-dimensional operation data of the optical storage and charging system are acquired through sensor network, and the acquired multi-dimensional operation data are preprocessed to obtain historical datasets and real-time datasets;

[0009] (2) AI model training and prediction: The AI ​​joint prediction model is trained using the historical dataset, and the real-time dataset is processed by the trained AI joint prediction model to obtain the prediction results of photovoltaic power output and charging load demand.

[0010] (3) Constraint input: Input the constraint parameters during the operation of the photovoltaic energy storage and charging system to define the boundary range of power optimization allocation;

[0011] (4) Global power optimization allocation: Combining the prediction results obtained in step (2) with the constraint parameters input in step (3), a multi-objective optimization algorithm is used to solve the power command baseline value of the specified power module in order to achieve the global optimal power allocation planning under multiple objectives;

[0012] (5) Real-time power correction and execution: Based on the hardware operating parameters and predicted values, and the deviation between the hardware operating parameters and real-time values, the power command baseline value obtained in step (4) is corrected in real time, the corrected power command is sent to the specified power module, the operating status of the specified power module is monitored in real time, and the fault protection mechanism is triggered.

[0013] (6) Model adaptive update: continuously monitor the deviation between the predicted value and the real-time value, and adaptively update the parameters of the AI ​​joint prediction model based on the deviation feedback, so that the AI ​​joint prediction model can adapt to the dynamic changes of the environment and load.

[0014] As a preferred scheme for the dynamic power allocation control method of photovoltaic-storage-charging based on AI prediction, in step (1), the multi-dimensional operating data includes photovoltaic side data, energy storage side data, charging side data, grid side data and environmental side data;

[0015] The photovoltaic-side data includes the irradiance of the photovoltaic array, module temperature, output power, and MPPT operating status.

[0016] The energy storage side data includes energy storage SOC, charge / discharge current, single cell voltage, and battery temperature;

[0017] The charging-side data shown includes charging pile connection status, charging power demand, charging time, and user charging habit tags.

[0018] The grid-side data includes grid voltage, frequency, peak and off-peak electricity price periods, and grid-connected power restriction instructions;

[0019] The environmental data includes real-time weather data and future weather forecast data.

[0020] As a preferred scheme for the dynamic power allocation control method of photovoltaic-storage-charging based on AI prediction, step (1) includes preprocessing the collected multi-dimensional operating data, including:

[0021] The 3σ criterion is used to eliminate abnormal data caused by sensor malfunctions. The 3σ criterion satisfies... ,in, For a single data value, The mean of the dataset. The standard deviation of the dataset;

[0022] Missing data is filled using linear interpolation. The linear interpolation formula is as follows:

[0023]

[0024] In the formula, , Given the coordinates of the data points, , The coordinates of adjacent known data points The x-coordinate corresponds to the missing data. The result is the interpolation result.

[0025] Normalize the data to the range of 0 to 1 using the following formula:

[0026]

[0027] In the formula, The original data, The minimum value of the data. For the maximum value of the data, This is the normalized data.

[0028] As a preferred scheme for the dynamic power allocation control method of photovoltaic storage and charging based on AI prediction, in step (2), the AI ​​joint prediction model is a multi-input multi-output model based on LSTM and attention mechanism;

[0029] The input features of the AI ​​joint prediction model include historical photovoltaic power output data, historical charging load data, real-time environmental data, weather forecast data for a future set period, time period characteristics, and set user charging habit tags;

[0030] The AI ​​joint prediction model structure includes an input layer, three LSTM hidden layers, an attention layer, and an output layer. The input layer converts the feature parameters into a tensor with a dimension of 23×T, where T is the time step.

[0031] Each hidden layer of the LSTM has 128 neurons and is used to extract temporal correlations of features. The LSTM unit state update formula is:

[0032]

[0033] In the formula, This represents the current cell state. Output for the forget gate. This refers to the cell state at the previous time step. For input gate output, The candidate cell state is... Multiplication of elements;

[0034] The attention layer assigns weights to the feature vectors output by the LSTM layer. The weights are calculated using the following formula:

[0035]

[0036] In the formula, Let be the attention weight for the t-th feature. Let T be the attention score for the t-th feature, and T be the length of the feature sequence.

[0037] The output layer outputs the photovoltaic power output prediction curve for a future set time period. With charging load demand forecast curve The prediction error satisfies:

[0038]

[0039] In the formula, This is the actual value. For predicted values, The set prediction error deviation rate threshold.

[0040] As a preferred scheme for the dynamic power allocation control method of photovoltaic-storage-charging based on AI prediction, the constraints in step (3) include:

[0041] Energy storage SOC constraints:

[0042]

[0043] In the formula, For real-time state of charge of energy storage, This is the minimum state of charge for energy storage. This represents the maximum state of charge of the energy storage.

[0044] Grid-connected power constraints:

[0045]

[0046] In the formula, Power command exchange with the power grid. The upper limit of grid-connected power is determined by grid dispatch instructions;

[0047] Power balance constraints:

[0048]

[0049] In the formula, For photovoltaic output power command, This is the energy storage charging / discharging power command; charging is positive, discharging is negative. Power command required by charging load.

[0050] As a preferred scheme for the dynamic power allocation control method of photovoltaic-storage-charging based on AI prediction, in step (4), the multi-objective optimization algorithm is a particle swarm optimization algorithm, and the objective function expression is:

[0051] ;

[0052] In the formula, , , These are weighting coefficients, which are adjusted according to scenario requirements using the analytic hierarchy process. For photovoltaic power absorption rate, , This represents the actual power absorbed by the photovoltaic system. Forecast value of photovoltaic power output; The optimal state of charge for energy storage is used to balance charging and discharging capabilities with lifespan. For peak-valley electricity price arbitrage profits, , This refers to the power output from the power grid during off-peak hours. Off-peak electricity pricing This refers to the power output taken from the power grid during peak hours. This refers to peak hour electricity pricing.

[0053] As a preferred scheme for the dynamic power allocation control method of photovoltaic storage and charging based on AI prediction, in step (5), the power command reference value obtained in step (4) is corrected in real time, including hardware parameter correction and deviation correction.

[0054] The hardware parameter correction process involves acquiring the DC bus voltage. With module power transmission loss If the DC bus voltage If the voltage drops below the set lower limit, increase the energy storage discharge power or the power drawn from the grid; if the DC bus voltage... If the power exceeds the set upper limit, the energy storage charging power will be increased or the photovoltaic output power will be reduced.

[0055] The deviation correction process, if Then, based on the real-time photovoltaic output, the energy storage and grid power commands are readjusted, where, For real-time photovoltaic power output, Forecast values ​​of photovoltaic power output The set deviation rate threshold.

[0056] As a preferred scheme for the dynamic power allocation control method of photovoltaic-storage-charging based on AI prediction, in step (6), during the adaptive update of the parameters of the AI ​​joint prediction model according to the deviation feedback, the triggering condition for the adaptive update is: when the deviation rate is established for three consecutive time periods, the model parameters are updated using the gradient descent method, and the parameter update formula of the gradient descent method is:

[0057]

[0058] In the formula, For model parameters, For learning rate, This represents the gradient of the loss function with respect to the parameters.

[0059] This invention also provides an AI-predictive dynamic power allocation control system for photovoltaic-storage-charging systems, employing the aforementioned AI-predictive dynamic power allocation control method for photovoltaic-storage-charging systems, comprising:

[0060] The data acquisition and preprocessing unit is used to acquire multi-dimensional operating data of the photovoltaic storage and charging system through a sensor network, and to preprocess the acquired multi-dimensional operating data to obtain historical datasets and real-time datasets.

[0061] The AI ​​model training and prediction unit is used to train the AI ​​joint prediction model using the historical dataset, and to process the real-time dataset using the trained AI joint prediction model to obtain the prediction results of photovoltaic power output and charging load demand.

[0062] The constraint input unit is used to input the constraint parameters during the operation of the optical energy storage and charging system, and to define the boundary range of power optimization allocation;

[0063] The global power optimization allocation unit combines the prediction results obtained from the AI ​​model training and prediction unit with the constraint parameters input by the constraint input unit, and uses a multi-objective optimization algorithm to solve for the power command baseline value of the specified power module, so as to achieve the global optimal power allocation planning under multiple objectives.

[0064] The real-time power correction and execution unit is used to correct the power command baseline value obtained by the global power optimization allocation unit in real time based on the hardware operating parameters and predicted values, as well as the deviation between the hardware operating parameters and real-time values. The corrected power command is then sent to the designated power module, the operating status of the designated power module is monitored in real time, and the fault protection mechanism is triggered.

[0065] The model adaptive update unit is used to continuously monitor the deviation between the predicted value and the real-time value, and adaptively update the parameters of the AI ​​joint prediction model based on the deviation feedback, so that the AI ​​joint prediction model can adapt to the dynamic changes of the environment and load.

[0066] The present invention also provides a computer storage medium storing program code for a dynamic power allocation control method for optical storage and charging based on AI prediction, the program code including instructions for executing the above-described dynamic power allocation control method for optical storage and charging based on AI prediction.

[0067] The present invention has the following advantages:

[0068] First, this invention uses a multi-dimensional AI joint prediction model to predict the changing trends of photovoltaic power output and charging load, and combines a global optimization strategy to rationally allocate power, thereby reducing the curtailment of photovoltaic power caused by overcapacity and fully tapping the energy utilization potential of distributed photovoltaics.

[0069] Secondly, this invention constructs a scientific energy storage charging and discharging control mechanism, strictly controls the energy storage state of charge within a reasonable range, avoids damage to the battery from deep charging and discharging, and optimizes the charging and discharging frequency and power to reduce the cycle loss of the energy storage battery and extend the service life of the energy storage system.

[0070] Third, by taking advantage of the peak-valley electricity price difference, this invention optimizes the power allocation strategy to achieve an operation mode of energy storage charging during off-peak hours and energy storage discharging to supplement the load during peak hours, thereby reducing the amount of electricity drawn from the grid during peak hours and lowering electricity costs; at the same time, it reduces energy transmission losses and further reduces the overall operating cost of the system.

[0071] Fourth, this invention establishes a hardware-control collaborative adaptation mechanism, incorporating hardware operating parameters into the control logic to correct power commands in real time, avoiding problems such as bus voltage exceeding limits and module overload, and ensuring the stable operation of all system components. By controlling grid-connected power, it avoids large power fluctuations and backfeed phenomena, reduces the impact on the voltage and frequency of the public power grid, achieves friendly interaction with the power grid, and ensures the safe operation of the power grid.

[0072] Fifth, it supports users to customize the priority of control targets according to different application scenarios, flexibly adjust the weight of targets such as photovoltaic consumption, energy storage protection, charging demand, and economic efficiency, and adapt to the differentiated operation needs of various scenarios such as residential, commercial supercharging stations, and industrial parks.

[0073] Sixth, this invention dynamically balances the power supply of photovoltaics, energy storage, and the power grid, ensuring that the charging pile can obtain stable power support under different operating conditions, avoiding prolonged charging time due to insufficient power, and improving the user's charging experience.

[0074] Seventh, the present invention has an adaptive update mechanism, which can dynamically adjust the model parameters according to environmental changes, load characteristics changes, etc., to ensure the prediction accuracy and long-term effectiveness of the control strategy, and avoid the performance degradation problem of traditional models after long-term operation. Attached Figure Description

[0075] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0076] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0077] Figure 1 This is a schematic diagram of the dynamic power allocation control method for photovoltaic storage and charging based on AI prediction provided in an embodiment of the present invention;

[0078] Figure 2 This is a technical roadmap for the AI-predictive-based dynamic power allocation control method for photovoltaic storage and charging provided in this embodiment of the invention.

[0079] Figure 3 This is a diagram of the AI ​​joint prediction model architecture of the dynamic power allocation control method for photovoltaic storage and charging based on AI prediction provided in this embodiment of the invention.

[0080] Figure 4 This is an application architecture diagram of the dynamic power allocation control method for photovoltaic storage and charging based on AI prediction provided in the embodiments of the present invention;

[0081] Figure 5 This is a schematic diagram of the dynamic power allocation control system architecture for photovoltaic storage and charging based on AI prediction provided in an embodiment of the present invention. Detailed Implementation

[0082] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0083] Example 1

[0084] See Figure 1 and Figure 2 This invention provides a dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction, comprising the following steps:

[0085] S1. Data Acquisition and Preprocessing: Multi-dimensional operational data of the photovoltaic energy storage and charging system are acquired through a sensor network. The acquired multi-dimensional operational data is preprocessed to obtain historical datasets and real-time datasets.

[0086] The power allocation decisions of the photovoltaic-storage-charging system rely on comprehensive and high-quality data support. Multi-dimensional operational data covers the source, storage, load, grid, and environment to ensure data integrity. Data preprocessing eliminates noise interference such as anomalies and missing data in the raw data. Anomalies may be caused by sensor failures, electromagnetic interference, etc. The 3σ criterion ensures data reliability by removing data that deviates from the mean by more than three standard deviations. Linear interpolation can fill in missing values ​​caused by transmission delays and temporary sensor malfunctions during data acquisition, maintaining the continuity of data sequence. Normalization maps data of different dimensions and numerical ranges to the 0~1 interval, avoiding the impact of differences in data magnitude on the convergence speed and prediction accuracy of AI model training. Finally, the historical dataset is used for the model to learn data patterns, and the real-time dataset is used for online prediction.

[0087] S2. AI Model Training and Prediction: The historical dataset is used to train an AI joint prediction model. The trained AI joint prediction model is then used to process the real-time dataset to obtain the prediction results of photovoltaic power output and charging load demand.

[0088] The AI ​​joint prediction model employs a multi-input multi-output architecture using LSTM and an attention mechanism. Addressing the intermittency of photovoltaic power output and the randomness of charging load, it improves prediction accuracy by fusing multi-dimensional features. LSTM (Long Short-Term Memory) networks possess the ability to capture long-term dependencies in time-series data, effectively learning the patterns of photovoltaic power output and charging load changes over time. The configuration of 3 hidden layers and 128 neurons fully extracts deep features from the data. The attention mechanism assigns weights to the feature vectors output by the LSTM, strengthening the influence of key features such as irradiance and user charging periods while weakening interference from irrelevant features. During model training, parameters are iteratively optimized using historical datasets to establish a mapping relationship between features and prediction targets. During prediction, real-time data is input, and the trained model outputs photovoltaic power output and charging load curves for future periods, providing forward-looking decision-making basis for global power optimization.

[0089] S3. Constraint Input: Input the constraint parameters during the operation of the photovoltaic energy storage and charging system to define the boundary range of power optimization allocation.

[0090] Among them, the energy storage SOC constraint limits the upper and lower limits of the state of charge to avoid irreversible damage to the battery caused by deep charging and discharging, thus extending the life of energy storage; the grid connection power constraint is determined by the grid dispatch command to prevent the system from sending excessive power to the grid, which would cause voltage and frequency fluctuations and ensure grid friendliness; the power balance constraint is a basic requirement of energy conservation, ensuring that the sum of the output power of photovoltaic, energy storage and grid matches the charging load demand, and avoiding abnormal system operation caused by power excess or deficiency.

[0091] S4. Global Power Optimization Allocation: Combining the prediction results obtained in step S2 with the constraint parameters input in step S3, a multi-objective optimization algorithm is used to solve for the power command baseline value of the specified power module, so as to achieve the global optimal power allocation planning under multiple objectives.

[0092] The core of global power optimization allocation lies in finding the optimal power allocation scheme within the constraint boundaries using an algorithm. Particle swarm optimization (PSO) is suitable for multi-objective optimization scenarios due to its fast convergence speed and robustness. It iteratively finds the maximum value of the objective function by simulating the search behavior of particles in the solution space. The objective function integrates photovoltaic (PV) grid integration rate, energy storage lifetime protection, and operational economic objectives. The weight coefficients are adjusted according to scenario requirements using the analytic hierarchy process (AHP) to adapt the priority of objectives to different scenarios. During the algorithm's solution process, AI prediction results are used as input, and feasible solutions are selected based on the constraints. The final output is a baseline value for power commands from PV, energy storage, and the grid. This ensures sufficient PV energy utilization, reduces energy storage losses, and lowers operating costs through peak-valley electricity price arbitrage, achieving global optimization across multiple objectives.

[0093] S5. Real-time power correction and execution: Based on the hardware operating parameters and predicted values, as well as the deviation between the hardware operating parameters and real-time values, the power command baseline value obtained in step S4 is corrected in real time, the corrected power command is sent to the designated power module, the operating status of the designated power module is monitored in real time, and the fault protection mechanism is triggered.

[0094] Real-time power correction compensates for the impact of prediction deviations and dynamic hardware changes, ensuring the real-time performance and accuracy of power allocation. Hardware parameter correction addresses the characteristics of the common DC bus topology. The DC bus voltage is a key indicator of system stability. When the voltage falls below the lower limit or rises above the upper limit, the system adjusts the energy storage charging and discharging power, grid power intake, or photovoltaic output power to maintain the bus voltage within the set range, preventing module overload and voltage exceeding limits. Deviation correction addresses the difference between predicted and real-time values. When the photovoltaic output deviation exceeds a threshold, the system promptly adjusts the power commands for energy storage and the grid to ensure power balance. After commands are issued, the system monitors the operating status of each module in real time. If overload, overtemperature, or other faults occur, the protection mechanism is immediately triggered to disconnect the faulty module or adjust its power, ensuring the safe and stable operation of the system equipment.

[0095] S6. Model Adaptive Update: Continuously monitor the deviation between the predicted value and the real-time value, and adaptively update the parameters of the AI ​​joint prediction model based on the deviation feedback, so that the AI ​​joint prediction model can adapt to the dynamic changes of the environment and load.

[0096] Among these, adaptive model updates are key to overcoming the limitations of traditional "one-time training, lifelong use" models, ensuring the predictive accuracy of the model over long-term operation. During system operation, environmental conditions (such as changes in climate patterns) and load characteristics (such as changes in user charging habits) may dynamically change, leading to increased prediction bias in the original model. By continuously monitoring the deviation rate between predicted and real-time values, a model update mechanism is triggered when the deviation exceeds a threshold for multiple consecutive time periods. Gradient descent calculates the gradient of the loss function with respect to the model parameters and iteratively adjusts the parameters along the negative gradient direction, enabling the model to continuously learn new operating rules, adapt to dynamically changing scenarios, maintain good predictive performance, and ensure the long-term effectiveness of the power allocation strategy.

[0097] In one possible embodiment, in step S1, the multi-dimensional operational data includes photovoltaic-side data, energy storage-side data, charging-side data, grid-side data, and environmental-side data; the photovoltaic-side data includes the irradiance of the photovoltaic array, module temperature, output power, and MPPT operating status; the energy storage-side data includes energy storage SOC, charging and discharging current, single-cell voltage, and battery temperature; the charging-side data includes the charging pile connection status, charging power demand, charging duration, and user charging habit tags; the grid-side data includes grid voltage, frequency, peak and off-peak electricity price periods, and grid-connected power limitation instructions; and the environmental-side data includes real-time weather data and future weather forecast data.

[0098] Specifically, the selection of multi-dimensional operational data ensures that the data fully reflects the operational status and influencing factors of each link in the system. Photovoltaic (PV) data directly determines the PV output potential; irradiance and module temperature are core influencing factors for PV power, and the MPPT (Multi-Level Photovoltaic Power Test) status reflects the operating efficiency of PV modules. Energy storage data is used to assess the remaining capacity, health status, and charging / discharging capabilities of energy storage, providing a basis for energy storage charging and discharging control. Charging data captures user charging behavior characteristics; user charging habit tags (such as commuters and logistics users) can help predict load distribution patterns. Grid-side data is linked to grid operation status and electricity pricing policies, and is key to achieving grid-friendly interaction and economic optimization. Environmental data (such as temperature, wind speed, and weather forecasts) is used to predict PV output trends, improving the foresight of prediction models. The integration of multi-dimensional data avoids the limitations of a single data dimension.

[0099] In one possible embodiment, step S1, preprocessing the collected multi-dimensional operational data, includes:

[0100] The 3σ criterion is used to eliminate abnormal data caused by sensor malfunctions. The 3σ criterion satisfies... ,in, For a single data value, The mean of the dataset. Let μ be the standard deviation of the dataset. The 3σ criterion is based on the characteristics of the normal distribution. In a normal distribution, approximately 99.73% of the data falls within the mean ± 3σ range. Data outside this range are considered outliers, likely caused by abnormal factors such as sensor malfunctions or data transmission errors. By calculating the mean μ and standard deviation σ of the dataset, data that satisfy |x-μ|>3σ are filtered out and removed, effectively eliminating noise interference and ensuring data reliability.

[0101] Among them, linear interpolation is used to complete the missing data. The linear interpolation formula is as follows:

[0102]

[0103] In the formula, , Given the coordinates of the data points, , The coordinates of adjacent known data points The x-coordinate corresponds to the missing data. This is the interpolation result. Linear interpolation assumes a linear relationship between two adjacent known data points, making it suitable for scenarios with small amounts of missing data and stable trends. It is a simple and efficient method for completing missing values ​​in engineering applications. By establishing a linear function relationship using the coordinates of adjacent known data points, the corresponding value of the missing data point can be calculated, maintaining the temporal continuity and trend consistency of the data.

[0104] The data is normalized to the 0-1 range using the following formula:

[0105]

[0106] In the formula, The original data, The minimum value of the data. For the maximum value of the data, This is normalized data. Data from different dimensions varies significantly in terms of units and numerical ranges. Directly inputting this data into the model can lead to features with large numerical values ​​dominating, while features with small numerical values ​​are ignored, negatively impacting model training performance. Normalization maps all data to the 0-1 range through a linear transformation, eliminating differences in units and numerical ranges and balancing the influence of each feature on the model. Furthermore, the normalized numerical range better matches the input requirements of AI models, accelerating model training convergence and improving model stability and prediction accuracy.

[0107] See Figure 3 In one possible embodiment, in step S2, the AI ​​joint prediction model is a multi-input multi-output model based on LSTM and attention mechanism;

[0108] The input features of the AI ​​joint prediction model include historical photovoltaic (PV) output data, historical charging load data, real-time environmental data, weather forecast data for a future set time period, time-period characteristics, and user charging habit tags. Historical PV output data and historical charging load data form the basis for the model to learn temporal variation patterns, reflecting intraday and interday trends. Real-time environmental data, such as irradiance and temperature, directly affect current PV output. Weather forecast data for a future set time period is used to predict future trends in PV output. Time-period characteristics and user charging habit tags capture the temporal distribution patterns of charging load. The fusion of multi-dimensional input features avoids the limitations of single features, provides rich information support for the model, and improves prediction accuracy.

[0109] The AI ​​joint prediction model structure includes an input layer, three LSTM hidden layers, an attention layer, and an output layer. The input layer converts the feature parameters into a 23×T tensor, where T is the time step. The model structure is designed to adapt to the processing requirements of multi-dimensional time-series data. The input layer converts the 23 feature parameters into a 23×T tensor, which meets the input format requirements of the LSTM model, enabling simultaneous input of multi-feature time-series data. The configuration of three LSTM hidden layers can deeply extract the temporal dependencies of the data. A single hidden layer is difficult to capture complex temporal patterns, and the multi-layer structure can improve the model's feature extraction capability. The attention layer is used to strengthen the influence of key features, solving the problem that the LSTM model does not pay enough attention to key information in long-series data. The output layer, for the multi-objective prediction requirements, simultaneously outputs two prediction curves: photovoltaic power output and charging load, achieving dual-objective output with a single prediction, improving prediction efficiency, and adapting to the power allocation requirements of photovoltaic-storage-charging systems.

[0110] In this LSTM hidden layer, each layer has 128 neurons, used to extract temporal correlations of features. The LSTM unit state update formula is as follows:

[0111]

[0112] In the formula, This represents the current cell state. Output for the forget gate. This refers to the cell state at the previous time step. For input gate output, The candidate cell state is... Element-wise multiplication. Forget gate output. Control the state of the unit at the previous moment The retention ratio can discard irrelevant historical information; input gate output Control the current candidate cell state The input ratio is used to select important current information and store it in the cell state; element-wise multiplication. The weights are assigned element by element, ultimately yielding the current cell state. This approach retains key historical information while incorporating the latest input, enabling the model to effectively learn temporal relationships in long-sequence data, such as the lag relationship between irradiance changes and photovoltaic output. The 128-neuron configuration ensures the model has sufficient feature representation capabilities to capture complex temporal patterns.

[0113] The attention layer assigns weights to the feature vectors output by the LSTM layer, and the weight calculation formula is as follows:

[0114]

[0115] In the formula, Let be the attention weight for the t-th feature. Let T be the attention score for the t-th feature, where T is the length of the feature sequence. Attention Score The importance of the t-th feature to the prediction target is reflected by calculating the similarity between the LSTM output feature and the query vector. The weight calculation formula uses the softmax function to convert the attention score into a weight value in the range of 0 to 1, and the sum of the weights of all features is 1, achieving normalized weight allocation. For photovoltaic power output prediction, features such as irradiance and temperature will have higher weights; for charging load prediction, user charging habits and time period features will have higher weights. By assigning higher weights to key features, the model focuses on core influencing factors when calculating prediction results, weakening the interference of irrelevant features, thereby improving prediction accuracy.

[0116] The output layer outputs the photovoltaic power output prediction curve for a future set time period. With charging load demand forecast curve The prediction error satisfies:

[0117]

[0118] In the formula, This is the actual value. For predicted values, The set prediction error deviation rate threshold is defined. The output layer adopts a multi-output architecture, simultaneously outputting two prediction curves: photovoltaic power output and charging load. This adapts to the "source-load" collaborative prediction requirements of photovoltaic-storage-charging systems, and the time resolution can meet the decision-making frequency of real-time power allocation. The prediction error formula quantifies the model's prediction accuracy. By calculating the relative error between the actual and predicted values, it ensures that the error is controlled within the set threshold ε. Relative error, compared to absolute error, better reflects the rationality of prediction accuracy, avoiding distortion in error assessment caused by differences in numerical magnitude. Strict error control ensures that the prediction results provide a reliable basis for global power optimization and allocation, preventing the power allocation strategy from failing due to excessive prediction deviation.

[0119] In one possible embodiment, in step S3, the constraints include:

[0120] Energy storage SOC constraints:

[0121]

[0122] In the formula, For real-time state of charge of energy storage, This is the minimum state of charge for energy storage. This represents the maximum state of charge (SOC) for energy storage. The SOC constraint is a safe operating boundary set based on battery characteristics. The SOC of an energy storage battery directly affects its lifespan and safety performance. An SOC below the minimum value leads to deep discharge, accelerating the degradation of active materials inside the battery and shortening cycle life; an SOC above the maximum value leads to overcharging, causing safety risks such as battery overheating and bulging, while also reducing battery capacity retention. By constraining the real-time SOC of the energy storage battery between a certain range, deep charging / discharging and overcharging / over-discharging can be avoided, ensuring the safe and stable operation of the energy storage battery and extending its service life.

[0123] Grid-connected power constraints:

[0124]

[0125] In the formula, Power command exchange with the power grid. The upper limit of grid-connected power is determined by grid dispatch instructions. Grid-connected power constraints are crucial for ensuring grid friendliness, as the public grid's capacity is limited. If the power transmitted from the photovoltaic-storage-charging system to the grid exceeds [a certain limit], [further constraints will be imposed]. (Determined by the power grid dispatching system based on factors such as line capacity and load levels), this can lead to increased grid voltage and frequency fluctuations, affecting the safe and stable operation of the power grid and even triggering the operation of power grid protection devices. Limiting the power grid's interactive power commands can address this issue. No more than This can avoid excessive power backfeeding from impacting the power grid, ensure friendly interaction between the system and the power grid, and meet the requirements of power grid dispatch.

[0126] Power balance constraints:

[0127]

[0128] In the formula, For photovoltaic output power command, This is the energy storage charging / discharging power command; charging is positive, discharging is negative. This refers to the power command required by the charging load. Power balance constraints, adhering to the law of conservation of energy, are a fundamental requirement for the stable operation of photovoltaic-storage-charging systems. Photovoltaic output power. Energy storage charging / discharging power (When discharge is negative, it is equivalent to output power), grid interaction power. (When power is drawn, it is equivalent to input power) together constitute the energy supply side of the system, and the charging load requires power. This is the energy consumption side. If the power supplied exceeds the power demanded, the DC bus voltage will rise, damaging the power module; if the power supplied is less than the power demanded, the DC bus voltage will drop, preventing the charging equipment from operating normally. Power balance constraints ensure real-time matching between the power supplied and demanded, maintaining system energy balance and guaranteeing stable equipment operation.

[0129] In one possible embodiment, in step S4, the multi-objective optimization algorithm is a particle swarm optimization algorithm, and the objective function expression is:

[0130] ;

[0131] In the formula, , , These are weighting coefficients, which are adjusted according to scenario requirements using the analytic hierarchy process. For photovoltaic power absorption rate, , This represents the actual power absorbed by the photovoltaic system. Forecast value of photovoltaic power output; The optimal state of charge for energy storage is used to balance charging and discharging capabilities with lifespan. For peak-valley electricity price arbitrage profits, , This refers to the power output from the power grid during off-peak hours. Off-peak electricity pricing This refers to the power output taken from the power grid during peak hours. This refers to peak hour electricity pricing.

[0132] Specifically, the objective function is designed to balance the goals of photovoltaic power consumption, energy storage lifespan, and operational economics. The particle swarm optimization algorithm can efficiently solve this multi-objective optimization problem. (Weight coefficients) , , Determined using the analytic hierarchy process (AHP), the system can be flexibly adjusted according to scenario requirements; for example, commercial supercharging stations prioritize cost-effectiveness. Higher weighting; residential systems focus on photovoltaic (PV) integration. Higher weighting ensures better alignment with target priorities. Photovoltaic grid integration rate. Reflects the utilization level of photovoltaic energy and maximizes its effectiveness. This can reduce wasted light.

[0133] in, This item reflects the degree of deviation between the energy storage state and the optimal state of charge. A larger value indicates that the energy storage is closer to its optimal operating state, which can reduce lifespan losses; peak-valley electricity price arbitrage income. Economic efficiency is improved by taking electricity at lower prices during off-peak hours and taking less electricity during peak hours (supplemented by energy storage discharge). The particle swarm optimization algorithm searches for the optimal solution by simulating particles, finding a power allocation scheme that maximizes the objective function under constraints, thus achieving multi-objective collaborative optimization.

[0134] In one possible embodiment, in step S5, the power command reference value obtained in step S4 is corrected in real time, including hardware parameter correction and deviation correction.

[0135] The hardware parameter correction process involves acquiring the DC bus voltage. With module power transmission loss If the DC bus voltage If the voltage drops below the set lower limit, increase the energy storage discharge power or the power drawn from the grid; if the DC bus voltage... If the voltage exceeds the set upper limit, the energy storage charging power will be increased or the photovoltaic output power will be reduced. Hardware parameter correction addresses the dynamic characteristics of the common DC bus topology, including the DC bus voltage. It is an indicator reflecting the system's energy balance and hardware operating status, including module power transmission loss. This will affect energy utilization efficiency and equipment heat generation. When If the voltage is below the set lower limit, it indicates insufficient energy supply. Increasing the energy storage discharge power or the power drawn from the grid can supplement energy and improve the bus voltage. If the voltage exceeds the set upper limit, it indicates an excess of energy supply. Increasing the energy storage charging power (absorbing excess energy) or reducing the photovoltaic output power (reducing energy supply) can lower the bus voltage. By correcting hardware parameters, the power command is adapted to the hardware operating status in real time, avoiding voltage exceeding limits and module overload, and ensuring stable system operation.

[0136] Wherein, the deviation correction process, if Then, based on the real-time photovoltaic output, the energy storage and grid power commands are readjusted, where, For real-time photovoltaic power output, Forecast values ​​of photovoltaic power output This is the set deviation rate threshold. Deviation correction is used to compensate for prediction errors in the AI ​​prediction model. Real-time photovoltaic output is affected by unforeseen factors such as cloud cover. Possibly related to the predicted value There is a deviation. When the deviation rate exceeds the threshold... If power is still allocated according to the original baseline value, it will lead to a disruption of power balance. For example, if the predicted value is too high and the actual output is insufficient, it will result in a shortage of power supply to the load. In this case, by readjusting the power commands of energy storage and the grid, the insufficient output of photovoltaics can be supplemented by energy storage discharge or grid power intake, or the excess output of photovoltaics can be absorbed by energy storage charging, thus maintaining power balance, ensuring that the charging load demand is met, and improving the robustness of system operation.

[0137] In one possible embodiment, in step S6, during the adaptive update of the parameters of the AI ​​joint prediction model based on the deviation feedback, the triggering condition for the adaptive update is: when the deviation rate is met for three consecutive time periods, the model parameters are updated using the gradient descent method, and the gradient descent parameter update formula is:

[0138]

[0139] In the formula, For model parameters, For learning rate, This represents the gradient of the loss function with respect to the parameters.

[0140] Specifically, the trigger condition for adaptive model updates is set to an excessive bias rate for three consecutive time periods. This avoids erroneous updates caused by accidental factors such as momentary cloud cover or temporary sensor fluctuations, ensuring the necessity and accuracy of the updates. Gradient descent iteratively adjusts parameters along the negative direction of the loss function's gradient to minimize the loss function. The loss function quantifies the deviation between predicted and actual values, and its gradient... The learning rate α reflects the degree of influence of parameter changes on the loss function; it controls the step size of parameter updates. Too large a step size can cause model oscillations and non-convergence, while too small a step size can lead to slow convergence. The model parameters θ are updated using gradient descent, allowing the model to continuously learn new operating rules, adapt to dynamic changes in the environment and load, and maintain long-term prediction accuracy.

[0141] See Figure 4 This is an application architecture of the method of the present invention, which constructs an efficient and flexible dynamic power allocation system for photovoltaic storage and charging through five levels of organic linkage.

[0142] The lowest hardware coupling layer adopts a common DC bus coupling topology design, where all power modules are directly connected to the DC bus, eliminating the multiple AC-DC / DC-AC conversion stages found in traditional AC bus architectures and significantly reducing energy transmission losses. This hardware coupling layer integrates photovoltaic power modules, energy storage power modules, charging power modules, grid interaction modules, and auxiliary control modules. The photovoltaic modules maximize output through MPPT (Multi-Level Photovoltaic Test), the energy storage modules provide charge / discharge protection and status acquisition capabilities, the charging modules support 0-60kW power regulation and can be expanded to supercharging levels, the grid interaction module enables bidirectional power flow and grid connection restrictions, and the auxiliary control module maintains the bus voltage stable within a set range using a PID controller while simultaneously acquiring power data at a frequency of 1 minute. Furthermore, each power module uses a standardized pluggable interface, supporting plug-and-play expansion. Adding new devices requires no modification to the existing topology, significantly improving system flexibility and maintainability.

[0143] The data perception layer is responsible for collecting and preprocessing operational data across all dimensions, providing high-quality data support for upper-level decision-making. The data collection scope comprehensively covers the photovoltaic (PV) side, energy storage side, charging side, grid side, and environment side. The PV side includes key parameters such as irradiance and module temperature. The energy storage side uses a combination of ampere-hour integration and internal resistance methods to detect State of Charge (SOC), achieving an accuracy of ±2%. The charging side not only collects data such as charging demand power but also labels user charging habits. The grid side covers information such as voltage and peak / valley electricity price periods. The environment side integrates real-time weather and 72-hour weather forecast data. The collected data undergoes a series of preprocessing steps through an edge computing gateway. First, the 3σ criterion is used to remove abnormal data caused by sensor malfunctions. Then, linear interpolation is used to complete missing data. Subsequently, power and other data are normalized to the 0-1 range. Finally, the data is divided into historical and real-time datasets for AI model training and online prediction, respectively, ensuring data reliability and consistency and avoiding noise interference in subsequent decision-making.

[0144] The AI ​​prediction layer is a multi-input multi-output joint prediction model based on LSTM and attention mechanisms, achieving coordinated and accurate prediction of photovoltaic power output and charging load. The model's input features are rich and comprehensive, covering historical photovoltaic power output and charging load data for the past 7 days, real-time environmental data, 72-hour weather forecast data, time-period features, and 10 categories of user charging habit tags, totaling 23 parameters across 6 categories. This effectively captures various key factors affecting photovoltaic power output and charging load. The model structure is progressively designed. The input layer converts the 23 feature parameters into a 23×144-dimensional tensor. Three LSTM hidden layers (128 neurons per layer) are responsible for extracting the temporal relationships between features. The attention layer assigns weights to the feature vectors output by the LSTM layers, giving higher weights to key features such as recent irradiance and user charging time periods, further improving prediction accuracy. The output layer ultimately generates photovoltaic power output prediction curves and charging load demand prediction curves for the next 1-24 hours, with a time resolution of 15 minutes per segment. The model was trained using a dataset from the past year, and the prediction error on the test set was only 7.2%. It also features an online adaptive update mechanism. When the deviation between the real-time data and the predicted data exceeds 10% for three consecutive time periods, the model parameters will be updated using gradient descent to ensure that the model can adapt to the dynamic changes in the environment and load.

[0145] The decision results from the AI ​​prediction layer are transmitted to the power control layer, which employs a hierarchical control strategy that balances global optimization and real-time adjustment to achieve multi-objective balance and hardware adaptation. The global optimization layer uses a multi-objective function as its core, considering photovoltaic (PV) grid integration rate, energy storage lifespan, and operational economics. Through the analytic hierarchy process (AHP), weighting coefficients can be adjusted according to different scenario requirements; for example, commercial supercharging stations can prioritize economics, while residential systems can prioritize PV grid integration. Simultaneously, energy storage SOC constraints, grid-connected power constraints, and power balance constraints are set to ensure the optimization process operates within safe boundaries. Then, a particle swarm optimization algorithm (population size 50, iterations 100) is used to solve for the optimal power allocation scheme, outputting the baseline power command values ​​for each module. The real-time adjustment layer dynamically corrects the baseline values. On one hand, it adjusts hardware parameters based on DC bus voltage and module power transmission losses to prevent bus voltage exceeding limits. On the other hand, it compares the deviation between predicted and real-time values. When the PV output deviation exceeds 10%, the energy storage and grid power commands are readjusted. The corrected power commands are then sent to each power module controller via industrial Ethernet to ensure accurate power allocation.

[0146] The human-machine interface layer provides visual monitoring and parameter configuration functions, enabling users to easily operate and manage the system. Through the operation status monitoring interface, users can view key data such as photovoltaic output, energy storage SOC, charging load, and grid interaction power in real time, intuitively understanding the power flow. The parameter configuration interface supports custom weighting coefficients, upper and lower limits of energy storage SOC, and grid-connected power limits to meet personalized needs. The target priority selection interface offers four preset modes: "PV absorption priority," "charging demand priority," "economic priority," and "balanced mode," which users can switch between with a single click. Furthermore, it can automatically generate daily / monthly / yearly operation reports, statistically analyzing key indicators such as PV absorption rate, energy storage cycle count, and electricity cost savings, providing users with comprehensive operational references. Data communication between different layers is achieved via industrial Ethernet based on the Modbus TCP / IP protocol, ensuring real-time and reliable information transmission. The entire architecture not only solves the technical pain points of traditional photovoltaic-energy storage-charging systems but also adapts to the needs of various scenarios such as residential, commercial supercharging stations, and industrial parks, forming a complete and efficient dynamic power allocation solution for photovoltaic-energy storage-charging systems.

[0147] To verify the technical effects of the present invention, three typical scenarios were selected: commercial supercharging stations, residential photovoltaic-storage-charging microgrids, and industrial parks. Through comparative tests between traditional solutions and the present invention, the verification was carried out from five indicators: photovoltaic absorption rate, energy storage life, operating cost, transmission loss, and grid impact. The test period for each scenario was one year.

[0148] See Table 1 for comparison results of commercial supercharging station scenarios (100kW photovoltaic, 200kWh energy storage, 4 x 60kW supercharging piles):

[0149] Table 1 Comparison Results of Commercial Supercharging Station Scenarios

[0150]

[0151] This invention uses AI to predict periods of low charging load (22:00-8:00 the next day), storing low-priced grid electricity (0.3 yuan / kWh) in energy storage, and discharging it during peak hours (8:00-22:00) to supplement the charging load, achieving an annual arbitrage profit of 187,000 yuan, while traditional solutions, lacking prediction capabilities, only achieve 32,000 yuan in arbitrage. During the test, there were 128 instances of sudden drops in photovoltaic power due to cloud cover. This invention, through real-time deviation correction, resulted in only 8,200 kWh of wasted solar power, compared to 32,000 kWh for traditional solutions. The photovoltaic absorption rate increased from 72% to 96.5%, meaning approximately 24.5% more solar energy resources can be utilized annually, directly translating into economic benefits. The energy storage degradation rate decreased by 3.2 percentage points, from 8.3% to 5.1%, significantly extending the lifespan of the energy storage system and reducing equipment replacement costs. Annual electricity costs are reduced by 214,000 yuan (24%), and through more efficient energy management, nearly a quarter of annual electricity expenses can be saved. Energy transmission losses are reduced by 8.4 percentage points, significantly reducing energy waste during transmission. Grid-connected power fluctuations are reduced by 14.3 percentage points: from ±18.5% to ±4.2%, significantly improving grid stability and reducing the impact on the grid.

[0152] See Table 2 for comparison results of a residential photovoltaic-storage-charging microgrid scenario (5kW photovoltaic, 10kWh energy storage, and one 7kW charging pile):

[0153] Table 2 Comparison Results of Household Photovoltaic Storage Charging Microgrid Scenarios

[0154]

[0155] This invention predicts users' commuting charging habits (charging between 18:00 and 22:00 in the evening), storing excess photovoltaic power in energy storage during the day and relying entirely on energy storage for power at night. This results in 286 days of zero grid dependence charging per year, compared to only 32 days with traditional solutions. The photovoltaic absorption rate jumps from 68% to 98.2%, achieving almost full utilization of solar energy resources. The energy storage cycle life is extended by 28%, from 820 cycles to 1050 cycles, significantly improving the lifespan of the energy storage system and reducing equipment replacement costs. Annual charging costs decrease from 2150 yuan to 186 yuan, saving nearly 90% of charging costs annually. Energy transmission loss decreases from 11.3% to 3.8%, significantly reducing energy waste during transmission. Peak grid demand decreases from 7.0kW to 1.2kW, significantly reducing peak demand on the grid and improving grid stability.

[0156] See Table 3 for the comparison results of the industrial park scenario (500kW photovoltaic, 1MWh energy storage, 20 30kW charging piles):

[0157] Table 3 Comparison Results of Industrial Park Scenarios

[0158]

[0159] This invention incorporates production load into an AI prediction model, enabling coordinated scheduling of photovoltaic output, production load, and charging load. The photovoltaic absorption rate jumped from 65% to 94.8%, an increase of 29.8 percentage points, significantly improving the utilization efficiency of solar energy resources. The annual energy storage degradation rate decreased from 9.1% to 5.7%, a reduction of 3.4 percentage points, extending the lifespan of the energy storage system and reducing equipment replacement costs. Annual electricity expenses decreased from 2.385 million yuan to 1.523 million yuan, a saving of 36%, or 862,000 yuan annually. Energy transmission loss decreased from 13.5% to 4.7%, a reduction of 8.8 percentage points, reducing energy waste during transmission. The peak-valley load difference in the power grid decreased from 420kW to 185kW, a reduction of 235kW, significantly reducing grid load fluctuations and improving grid stability.

[0160] Example 2

[0161] See Figure 5 Embodiment 2 of the present invention also provides a dynamic power allocation control system for photovoltaic-storage-charging based on AI prediction, employing the dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction described in the above embodiments, including:

[0162] The data acquisition and preprocessing unit 100 is used to acquire multi-dimensional operating data of the optical storage and charging system through a sensor network, and to preprocess the acquired multi-dimensional operating data to obtain historical datasets and real-time datasets.

[0163] AI model training and prediction unit 200 is used to train an AI joint prediction model using the historical dataset, and to process the real-time dataset using the trained AI joint prediction model to obtain prediction results of photovoltaic power output and charging load demand.

[0164] The constraint input unit 300 is used to input constraint parameters during the operation of the optical storage and charging system and to define the boundary range of power optimization allocation.

[0165] The global power optimization allocation unit 400 is used to combine the prediction results obtained by the AI ​​model training and prediction unit with the constraint parameters input by the constraint input unit, and use a multi-objective optimization algorithm to solve for the power command baseline value of the specified power module, so as to achieve the global optimal power allocation planning under multiple objectives.

[0166] The real-time power correction and execution unit 500 is used to correct the power command baseline value obtained by the global power optimization allocation unit in real time based on the hardware operating parameters and predicted values, as well as the deviation between the hardware operating parameters and real-time values. The corrected power command is sent to the designated power module, the operating status of the designated power module is monitored in real time, and the fault protection mechanism is triggered.

[0167] The model adaptive update unit 600 is used to continuously monitor the deviation between the predicted value and the real-time value, and adaptively update the parameters of the AI ​​joint prediction model according to the deviation feedback, so that the AI ​​joint prediction model can adapt to the dynamic changes of the environment and load.

[0168] In this embodiment, the multi-dimensional operating data in the data acquisition and preprocessing unit 100 includes photovoltaic side data, energy storage side data, charging side data, grid side data, and environmental side data;

[0169] The photovoltaic-side data includes the irradiance of the photovoltaic array, module temperature, output power, and MPPT operating status.

[0170] The energy storage side data includes energy storage SOC, charge / discharge current, single cell voltage, and battery temperature;

[0171] The charging-side data shown includes charging pile connection status, charging power demand, charging time, and user charging habit tags.

[0172] The grid-side data includes grid voltage, frequency, peak and off-peak electricity price periods, and grid-connected power restriction instructions;

[0173] The environmental data includes real-time weather data and future weather forecast data.

[0174] In this embodiment, the data acquisition and preprocessing unit 100 performs preprocessing on the acquired multi-dimensional operational data, including:

[0175] The 3σ criterion is used to eliminate abnormal data caused by sensor malfunctions. The 3σ criterion satisfies... ,in, For a single data value, The mean of the dataset. The standard deviation of the dataset;

[0176] Missing data is filled using linear interpolation. The linear interpolation formula is as follows:

[0177]

[0178] In the formula, , Given the coordinates of the data points, , The coordinates of adjacent known data points The x-coordinate corresponds to the missing data. The result is the interpolation result.

[0179] Normalize the data to the range of 0 to 1 using the following formula:

[0180]

[0181] In the formula, The original data, The minimum value of the data. For the maximum value of the data, This is the normalized data.

[0182] In this embodiment, in the AI ​​model training and prediction unit 200, the AI ​​joint prediction model is a multi-input multi-output model based on LSTM and attention mechanism.

[0183] The input features of the AI ​​joint prediction model include historical photovoltaic power output data, historical charging load data, real-time environmental data, weather forecast data for a future set period, time period characteristics, and set user charging habit tags;

[0184] The AI ​​joint prediction model structure includes an input layer, three LSTM hidden layers, an attention layer, and an output layer. The input layer converts the feature parameters into a tensor with a dimension of 23×T, where T is the time step.

[0185] Each hidden layer of the LSTM has 128 neurons and is used to extract temporal correlations of features. The LSTM unit state update formula is:

[0186]

[0187] In the formula, This represents the current cell state. Output for the forget gate. This refers to the cell state at the previous time step. For input gate output, The candidate cell state is... Multiplication of elements;

[0188] The attention layer assigns weights to the feature vectors output by the LSTM layer. The weights are calculated using the following formula:

[0189]

[0190] In the formula, Let be the attention weight for the t-th feature. Let T be the attention score for the t-th feature, and T be the length of the feature sequence.

[0191] The output layer outputs the photovoltaic power output prediction curve for a future set time period. With charging load demand forecast curve The prediction error satisfies:

[0192]

[0193] In the formula, This is the actual value. For predicted values, The set prediction error deviation rate threshold.

[0194] In this embodiment, the constraint conditions in the constraint input unit 300 include:

[0195] Energy storage SOC constraints:

[0196]

[0197] In the formula, For real-time state of charge of energy storage, This is the minimum state of charge for energy storage. This represents the maximum state of charge of the energy storage.

[0198] Grid-connected power constraints:

[0199]

[0200] In the formula, Power command exchange with the power grid. The upper limit of grid-connected power is determined by grid dispatch instructions;

[0201] Power balance constraints:

[0202]

[0203] In the formula, For photovoltaic output power command, This is the energy storage charging / discharging power command; charging is positive, discharging is negative. Power command required by charging load.

[0204] In this embodiment, the multi-objective optimization algorithm in the global power optimization allocation unit 400 is a particle swarm optimization algorithm, and the objective function expression is:

[0205] ;

[0206] In the formula, , , These are weighting coefficients, which are adjusted according to scenario requirements using the analytic hierarchy process. For photovoltaic power absorption rate, , This represents the actual power absorbed by the photovoltaic system. Forecast value of photovoltaic power output; The optimal state of charge for energy storage is used to balance charging and discharging capabilities with lifespan. For peak-valley electricity price arbitrage profits, , This refers to the power output from the power grid during off-peak hours. Off-peak electricity pricing This refers to the power output taken from the power grid during peak hours. This refers to peak hour electricity pricing.

[0207] In this embodiment, the real-time power correction and execution unit 500 performs real-time correction on the power command reference value obtained by the global power optimization allocation unit 400, including hardware parameter correction and deviation correction.

[0208] The hardware parameter correction process involves acquiring the DC bus voltage. With module power transmission loss If the DC bus voltage If the voltage drops below the set lower limit, increase the energy storage discharge power or the power drawn from the grid; if the DC bus voltage... If the power exceeds the set upper limit, the energy storage charging power will be increased or the photovoltaic output power will be reduced.

[0209] The deviation correction process, if Then, based on the real-time photovoltaic output, the energy storage and grid power commands are readjusted, where, For real-time photovoltaic power output, Forecast values ​​of photovoltaic power output The set deviation rate threshold.

[0210] In this embodiment, in the model adaptive update unit 600, during the process of adaptively updating the parameters of the AI ​​joint prediction model based on the deviation feedback, the triggering condition for adaptive update is: when the deviation rate is met for three consecutive time periods, the model parameters are updated using the gradient descent method. The gradient descent method parameter update formula is:

[0211]

[0212] In the formula, For model parameters, For learning rate, This represents the gradient of the loss function with respect to the parameters.

[0213] It should be noted that the information interaction and execution process between the various units of the above system are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0214] Example 3

[0215] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code of an AI-predicted dynamic power allocation control method for optical storage and charging. The program code includes instructions for executing the AI-predicted dynamic power allocation control method for optical storage and charging as described in Embodiment 1 or any possible implementation thereof.

[0216] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0217] Example 4

[0218] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0219] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute the AI-predictive dynamic power allocation control method for photovoltaic storage and charging based on Embodiment 1 or any possible implementation thereof by calling the program instructions.

[0220] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0221] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0222] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0223] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction, characterized in that, Includes the following steps: (1) Data acquisition and preprocessing: Multi-dimensional operation data of the optical storage and charging system are acquired through sensor network, and the acquired multi-dimensional operation data are preprocessed to obtain historical datasets and real-time datasets; (2) AI model training and prediction: The AI ​​joint prediction model is trained using the historical dataset, and the real-time dataset is processed by the trained AI joint prediction model to obtain the prediction results of photovoltaic power output and charging load demand. (3) Constraint input: Input the constraint parameters during the operation of the photovoltaic energy storage and charging system to define the boundary range of power optimization allocation; (4) Global power optimization allocation: Combining the prediction results obtained in step (2) with the constraint parameters input in step (3), a multi-objective optimization algorithm is used to solve the power command baseline value of the specified power module in order to achieve the global optimal power allocation planning under multiple objectives; (5) Real-time power correction and execution: Based on the hardware operating parameters and predicted values, and the deviation between the hardware operating parameters and real-time values, the power command baseline value obtained in step (4) is corrected in real time, the corrected power command is sent to the specified power module, the operating status of the specified power module is monitored in real time, and the fault protection mechanism is triggered. (6) Model adaptive update: continuously monitor the deviation between the predicted value and the real-time value, and adaptively update the parameters of the AI ​​joint prediction model according to the deviation feedback, so that the AI ​​joint prediction model can adapt to the dynamic changes of the environment and load; In step (4), the multi-objective optimization algorithm is the particle swarm optimization algorithm, and the objective function expression is: ； In the formula, , , These are weighting coefficients, which are adjusted according to scenario requirements using the analytic hierarchy process. For photovoltaic power absorption rate, , This represents the actual power absorbed by the photovoltaic system. Forecast value of photovoltaic power output; The optimal state of charge for energy storage is used to balance charging and discharging capabilities with lifespan. For peak-valley electricity price arbitrage profits, , This refers to the power output from the power grid during off-peak hours. Off-peak electricity pricing This refers to the power output taken from the power grid during peak hours. Peak hour electricity price; The power command baseline value obtained in step (4) is corrected in real time, including hardware parameter correction; The hardware parameter correction process involves acquiring the DC bus voltage. With module power transmission loss If the DC bus voltage If the voltage drops below the set lower limit, increase the energy storage discharge power or the power drawn from the grid; if the DC bus voltage... If the power exceeds the set upper limit, the energy storage charging power will be increased or the photovoltaic output power will be reduced.

2. The dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction according to claim 1, characterized in that, In step (1), the multi-dimensional operational data includes photovoltaic side data, energy storage side data, charging side data, grid side data and environmental side data; The photovoltaic-side data includes the irradiance of the photovoltaic array, module temperature, output power, and MPPT operating status. The energy storage side data includes energy storage SOC, charge / discharge current, single cell voltage, and battery temperature; The charging-side data shown includes charging pile connection status, charging power demand, charging time, and user charging habit tags. The grid-side data includes grid voltage, frequency, peak and off-peak electricity price periods, and grid-connected power restriction instructions; The environmental data includes real-time weather data and future weather forecast data.

3. The dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction according to claim 1, characterized in that, In step (1), the preprocessing of the collected multi-dimensional operational data includes: The 3σ criterion is used to eliminate abnormal data caused by sensor malfunctions. The 3σ criterion satisfies... ,in, For a single data value, The mean of the dataset. The standard deviation of the dataset; Missing data is filled using linear interpolation. The linear interpolation formula is as follows: ， In the formula, , Given the coordinates of the data points, , The coordinates of adjacent known data points The x-coordinate corresponds to the missing data. This is the interpolation result; Normalize the data to the range of 0 to 1 using the following formula: ， In the formula, The original data, The minimum value of the data. For the maximum value of the data, This is the normalized data.

4. The dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction according to claim 1, characterized in that, In step (2), the AI ​​joint prediction model is a multi-input multi-output model based on LSTM and attention mechanism; The input features of the AI ​​joint prediction model include historical photovoltaic power output data, historical charging load data, real-time environmental data, weather forecast data for a future set period, time period characteristics, and set user charging habit tags; The AI ​​joint prediction model structure includes an input layer, three LSTM hidden layers, an attention layer, and an output layer. The input layer converts the feature parameters into a tensor with a dimension of 23×T, where T is the time step. Each hidden layer of the LSTM has 128 neurons and is used to extract temporal correlations of features. The LSTM unit state update formula is: ， In the formula, This represents the current cell state. Output for the forget gate. This refers to the cell state at the previous time step. For input gate output, The candidate cell state is... Multiplication of elements; The attention layer assigns weights to the feature vectors output by the LSTM layer. The weights are calculated using the following formula: ， In the formula, Let be the attention weight for the t-th feature. Let T be the attention score for the t-th feature, and T be the length of the feature sequence. The output layer outputs the photovoltaic power output prediction curve for a future set time period. With charging load demand forecast curve The prediction error satisfies: ， In the formula, This is the actual value. For predicted values, The set prediction error deviation rate threshold.

5. The dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction according to claim 1, characterized in that, In step (3), the constraints include: Energy storage SOC constraints: ， In the formula, For real-time state of charge of energy storage, This is the minimum state of charge for energy storage. This represents the maximum state of charge of the energy storage. Grid-connected power constraints: ， In the formula, Power command exchange with the power grid. The upper limit of grid-connected power is determined by grid dispatch instructions; Power balance constraints: ， In the formula, For photovoltaic output power command, This is the energy storage charging / discharging power command; charging is positive, discharging is negative. Power command required by charging load.

6. The dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction according to claim 1, characterized in that, In step (5), the power command reference value obtained in step (4) is corrected in real time, including deviation correction; The deviation correction process, if Then, based on the real-time photovoltaic output, the energy storage and grid power commands are readjusted, where, For real-time photovoltaic power output, Forecast values ​​of photovoltaic power output The set deviation rate threshold.

7. The dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction according to claim 1, characterized in that, In step (6), during the adaptive update of the parameters of the AI ​​joint prediction model based on the deviation feedback, the triggering condition for the adaptive update is: when the deviation rate is met for three consecutive time periods, the model parameters are updated using the gradient descent method. The gradient descent parameter update formula is: ， In the formula, For model parameters, For learning rate, This represents the gradient of the loss function with respect to the parameters.

8. A dynamic power allocation control system for photovoltaic-storage-charging based on AI prediction, employing the dynamic power allocation control method for photovoltaic-storage-charging based on AI prediction as described in any one of claims 1 to 7, characterized in that, include: The data acquisition and preprocessing unit is used to acquire multi-dimensional operating data of the photovoltaic storage and charging system through a sensor network, and to preprocess the acquired multi-dimensional operating data to obtain historical datasets and real-time datasets. The AI ​​model training and prediction unit is used to train the AI ​​joint prediction model using the historical dataset, and to process the real-time dataset using the trained AI joint prediction model to obtain the prediction results of photovoltaic power output and charging load demand. The constraint input unit is used to input the constraint parameters during the operation of the optical energy storage and charging system, and to define the boundary range of power optimization allocation; The global power optimization allocation unit combines the prediction results obtained from the AI ​​model training and prediction unit with the constraint parameters input by the constraint input unit, and uses a multi-objective optimization algorithm to solve for the power command baseline value of the specified power module, so as to achieve the global optimal power allocation planning under multiple objectives. The real-time power correction and execution unit is used to correct the power command baseline value obtained by the global power optimization allocation unit in real time based on the hardware operating parameters and predicted values, as well as the deviation between the hardware operating parameters and real-time values. The corrected power command is then sent to the designated power module, the operating status of the designated power module is monitored in real time, and the fault protection mechanism is triggered. The model adaptive update unit is used to continuously monitor the deviation between the predicted value and the real-time value, and adaptively update the parameters of the AI ​​joint prediction model based on the deviation feedback, so that the AI ​​joint prediction model can adapt to the dynamic changes of the environment and load.

9. A computer storage medium, characterized in that, The computer storage medium stores program code for an AI-predictive dynamic power allocation control method for optical storage and charging, the program code including instructions for executing the AI-predictive dynamic power allocation control method for optical storage and charging as described in any one of claims 1 to 7.

Images