Virtual power plant load regulation method and system based on deep learning scheduling strategy

Abstract

The invention discloses a virtual power plant load regulation method and system based on a deep learning scheduling strategy, and relates to the technical field of power plant load regulation, and the method comprises the steps: obtaining multi-source heterogeneous data, and obtaining a feature tensor through preprocessing; the feature tensor is input into an LSTM-Transform hybrid network model, and a context code is output; inputting the context code into the time sequence convolutional network model, outputting an ultra-short-term prediction result, and further obtaining a short-term prediction result; constructing a multi-objective optimization model, and modeling constraint conditions; solving the multi-objective optimization model through a preset layered architecture to obtain an optimal scheduling strategy, and generating a scheduling instruction; and issuing the scheduling instruction to distributed resources in the virtual power plant, and executing and converting the scheduling instruction into an equipment action. The method solves the problems that in the prior art, the optimization target is single, the dynamic adaptive capacity is lacked, and the real-time fluctuation response speed of the power grid is limited to a certain extent.

Description

Technical Field

[0001] This invention relates to the field of power plant load regulation technology, and in particular to a virtual power plant load regulation method and system based on deep learning scheduling strategies. Background Technology

[0002] Virtual power plants (VPPs), as the core carrier integrating distributed energy resources, energy storage, and adjustable loads, need to achieve coordinated resource scheduling through dynamic load regulation to address challenges such as large fluctuations in renewable energy output, diversified user demands, and high grid stability requirements. Deep learning-based scheduling strategies, with their advantages in multi-source data processing and dynamic optimization, have become a key technological direction for load regulation in virtual power plants, supporting efficient operation in scenarios such as electricity market trading and zero-carbon industrial parks.

[0003] Virtual power plant (VPP) dynamic load regulation and control refers to the process of dynamically adjusting the adjustable load, energy storage system, and controllable power output within a VPP by integrating data prediction, multi-objective optimization, and real-time decision-making technologies, coordinating renewable energy consumption and grid interaction, and ultimately achieving optimal overall system efficiency. Its core requirements include accurately capturing renewable energy and load fluctuations, balancing multi-dimensional optimization objectives (such as cost, carbon emissions, and user satisfaction), and rapidly responding to real-time uncertainties.

[0004] Existing load resource regulation methods introduce optimal control theory and scheduling optimization algorithms, combined with a penalty term for power output changes, to optimize the regulation process of high-energy-consuming loads, avoid frequent fluctuations, and reduce load fluctuations by penalizing power changes, thereby achieving load regulation and improving grid stability. At the same time, a mixed integer programming algorithm is used to optimize the start-up and shutdown operations of load resources, ensuring the coordination between the switching status of load equipment and the regulation strategy. By combining parallel computing technology, the computational efficiency of the scheduling optimization process is improved, ensuring that the virtual power plant can adjust load resources in real time when the grid load fluctuates.

[0005] The study found that existing technologies only focus on the single objective of "minimizing dispatch costs," failing to cover the key needs of the current power system and lacking dynamic adaptability. Existing technologies do not address core objectives such as maximizing renewable energy consumption, minimizing carbon emissions, and maximizing user satisfaction, making them unsuitable for the multi-dimensional control needs of scenarios with high renewable energy penetration. Furthermore, the dispatch strategies of existing technologies rely on offline model solutions without incorporating online learning, resulting in limitations in response speed to real-time grid fluctuations. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a virtual power plant load regulation method and system based on deep learning scheduling strategies. This invention solves the problems of existing technologies having a single optimization objective, lacking dynamic adaptability, and having certain limitations in response speed to real-time fluctuations in the power grid.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a virtual power plant load regulation method based on a deep learning scheduling strategy, comprising: Acquire multi-source heterogeneous data and preprocess the multi-source heterogeneous data to obtain feature tensors; The feature tensor is input into a pre-trained LSTM-Transformer hybrid network model, and the context encoding is output. The context encoding is input into a pre-trained temporal convolutional network model, and local temporal fusion features are extracted to output ultra-short-term prediction results. The last time step feature in the local temporal fusion features is extracted and output short-term prediction results through a fully connected layer. A multi-objective optimization model is constructed, and the constraints are modeled. Based on short-term and ultra-short-term prediction results, the multi-objective optimization model is solved through a pre-defined hierarchical architecture that includes a short-term scheduling layer and a real-time control layer to obtain the optimal scheduling strategy and generate scheduling instructions. The scheduling instructions are sent to the distributed resources within the virtual power plant, executed, and transformed into equipment actions.

[0008] As a further technical solution, the multi-source heterogeneous data includes historical load data, new energy output data, meteorological data, electricity price and market data, and equipment status data; The preprocessing includes missing value imputation, outlier detection, dimensionality elimination, and feature alignment; wherein, missing value imputation uses linear interpolation or the nearest neighbor mean method, outlier detection uses interquartile range method, dimensionality elimination uses data standardization, and feature alignment aligns all data according to a unified timestamp.

[0009] As a further technical solution, the LSTM-Transformer hybrid network model includes an LSTM layer and a Transformer encoder. The feature tensor is sequentially passed through the forget gate, input gate, cell state update, and output gate of the LSTM layer to generate a hidden state sequence. The Transformer encoder receives the hidden state sequence, first generates a query vector, key vector, and value vector through linear projection, then calculates global feature association through a sub-attention mechanism, and finally outputs the context code through a position feedforward network.

[0010] As a further technical solution, the temporal convolutional network model includes a causal convolutional layer, a dilated convolutional layer, and an activation function layer. The context encoding is input into the pre-trained temporal convolutional network model. First, the causal convolutional layer ensures that the temporal sequence is reasonable. Then, the dilated convolutional layer expands the receptive field. Finally, the ReLU activation function outputs local temporal fusion features.

[0011] As a further technical solution, the multi-objective optimization model includes the following objectives: maximizing the absorption of new energy sources, minimizing operating costs, minimizing carbon emissions, minimizing grid fluctuations, and maximizing user satisfaction; specifically expressed as follows: ; ; ; ; ; in, To contribute to the predicted new energy sources, This refers to the actual amount of renewable energy power consumed. The electricity purchase price is t, and it is negative when the electricity is sold. and These represent the exchange power with the main network during time periods t and t-1, respectively. Let be the power generation cost coefficient of power source i. Let i be the output of power source i during time interval t. Let these represent the compensation costs for increasing and decreasing the load k, respectively. These represent the upward and downward adjustments of load k during time period t, respectively. Carbon intensity of grid electricity The carbon strength of power source i. All of these are comfort weights for load k.

[0012] As a further technical solution, the modeling of constraints includes a power balance constraint model, a new energy consumption constraint model, an adjustable load regulation constraint model, an energy storage dynamic constraint model, and a controllable power source constraint model; specifically represented as follows: ; ; ; ; ; ; ; ; ; ; in, Let i be the output of power source i during time interval t. This refers to the actual amount of renewable energy power consumed. To contribute to the predicted new energy sources, Let be the charging and discharging power of energy storage j during time period t, respectively. The exchange power with the main network during time period t. This is the total load demand forecast. These represent the upward and downward adjustments to load k during time period t, respectively. These represent the maximum allowable upward and downward adjustments to load k, respectively. Let k be the operating efficiency coefficient for load k. The total number of time periods in the adjustment cycle. The time step for a single period. Let k be the baseline power. Let k be the total energy demand of load k. The temperature-to-power conversion factor, For the lowest comfortable temperature, The initial temperature. For the highest comfortable temperature, , Let J represent the state of charge of energy storage j at time points t and t-1, respectively. The charging efficiency of energy storage j, Let j be the discharge efficiency of the energy storage j. For the rated capacity of energy storage j, , These represent the minimum and maximum permissible states of charge (SOC) of energy storage j, respectively. , These represent the initial and final states of charge of energy storage j, respectively. This represents the maximum permissible ramp rate for power supply i. , Let be the start / stop state variables of power supply i at times t and t-1, respectively. Let i be the continuous operating time of power supply i. Let i be the minimum continuous operating time of power supply i.

[0013] As a further technical solution, the specific method for solving the multi-objective optimization model using a pre-defined hierarchical architecture comprising a short-term scheduling layer and a real-time control layer is as follows: First, the hierarchical architecture includes a short-term scheduling layer and a real-time control layer. Then, based on the short-term prediction results, the short-term scheduling layer is used to solve the multi-objective optimization model using mixed-integer programming, outputting an hourly-level global scheduling strategy. Based on the ultra-short-term prediction results, the real-time control layer is used to perform deep reinforcement learning online decision-making on the multi-objective optimization model, outputting a minute-level scheduling strategy. Finally, based on the hourly-level global scheduling strategy and the minute-level scheduling strategy, the optimal scheduling strategy is obtained.

[0014] Secondly, the present invention provides a virtual power plant load regulation system based on a deep learning scheduling strategy, comprising: The data acquisition module is configured to: acquire multi-source heterogeneous data and preprocess the multi-source heterogeneous data to obtain feature tensors; The prediction module is configured to: input feature tensors into a pre-trained LSTM-Transformer hybrid network model and output context encoding; input the context encoding into a pre-trained temporal convolutional network model, extract local temporal fusion features, and output ultra-short-term prediction results; extract the last time step features from the local temporal fusion features and output short-term prediction results through a fully connected layer. The model building and solving module is configured to: build a multi-objective optimization model and model the constraints; based on short-term and ultra-short-term prediction results, solve the multi-objective optimization model through a preset hierarchical architecture to obtain the optimal scheduling strategy and generate scheduling instructions; The execution module is configured to: issue the scheduling instructions to distributed resources within the virtual power plant, execute them, and convert them into equipment actions.

[0015] One or more technical solutions of the present invention have the following beneficial effects: (1) This invention constructs a five-dimensional multi-objective optimization model that includes "maximizing new energy consumption, minimizing operating costs, minimizing carbon emissions, minimizing grid fluctuations, and maximizing user satisfaction," and models the constraints to enable this invention to adapt to the multi-dimensional regulation needs in scenarios with high new energy penetration. It breaks through the limitation of the single cost objective in existing technologies and can simultaneously adapt to diverse needs such as new energy consumption, low-carbon operation, and user experience.

[0016] (2) The present invention solves the multi-objective optimization model through a pre-defined hierarchical architecture including a short-term scheduling layer and a real-time control layer, and adopts a hybrid solution method combining MIP and DRL. Global optimization is achieved through MIP in the short-term scheduling layer, and online decision-making is achieved through DRL in the real-time control layer. This avoids the static decision-making defects of the offline model, significantly improves the anti-interference capability of the system, and effectively balances the optimization accuracy and response speed. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 This is a flowchart of the virtual power plant load regulation method based on deep learning scheduling strategy of the present invention. Detailed Implementation

[0019] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] Example 1 This embodiment provides a virtual power plant load regulation method based on a deep learning scheduling strategy, such as... Figure 1 As shown, the specific steps include the following: S1: Obtain multi-source heterogeneous data and preprocess the multi-source heterogeneous data to obtain feature tensors.

[0021] In step S1, the multi-source heterogeneous data includes historical load data read through the smart meter data interface, renewable energy output data collected through the wind and solar monitoring system API interface, meteorological data collected through meteorological station sensors, electricity price and market data collected through the power trading platform data interface, and equipment status data reported in real time through the EMS system.

[0022] In step S1, preprocessing includes missing value imputation, outlier detection, dimensionality elimination, and feature alignment. Missing value imputation employs either linear interpolation or a nearest neighbor method. Specifically, for time-series data, linear interpolation is used to imput missing values; for non-time-series data, the nearest neighbor method is used. The linear interpolation method is expressed as follows: ,in The interpolated value at time t, For the previous valid data point, For the next valid data point, k is the number of consecutive intervals for missing data.

[0023] Based on the nearest neighbor mean method, it can be expressed as follows: ;in The final imputation value for the i-th missing data point is the data used to fill the i-th missing position, calculated using the nearest neighbor mean method. The set of neighboring valid data points for the i-th missing data point refers to the set of all neighboring valid data points around the i-th missing position used to calculate the mean. The number of data points contained in the set of nearest valid data points N(i) is the total number of valid data points participating in the mean calculation. It corresponds to the denominator in the formula and is used to normalize the summation result. is the actual value of the j-th valid data point in the set of neighboring valid data points N(i). It is the basic data for calculating the mean and corresponds to the specific source of the summation term in the formula.

[0024] Outlier identification uses the interquartile range method, and the formula is as follows: ; ; in It is the first quartile (25th percentile). It is the third quartile (75th percentile).

[0025] Dimensional elimination employs data standardization to eliminate dimensional differences in the data, specifically using Z-Score standardization, with the formula: ,in The mean of the data. This represents the standard deviation of the data.

[0026] Feature alignment aligns all data according to a uniform timestamp.

[0027] In step S1, based on the preprocessed multi-source heterogeneous data, a feature tensor is obtained, which is used as input to the subsequent model. The feature tensor is... ,in For historical time steps, The feature dimensions include: 1) time series features with dimension D1; 2) time coding with dimension D2; 3) meteorological features with dimension D3; and 4) event features with dimension D4.

[0028] S2: Input the feature tensor into the pre-trained LSTM-Transformer hybrid network model and output the context encoding; input the context encoding into the pre-trained temporal convolutional network model, extract the local temporal fusion features, and output the ultra-short-term prediction results; extract the last time step features from the local temporal fusion features and output the short-term prediction results through the fully connected layer.

[0029] In step S2, to handle long temporal dependencies and multi-source feature fusion, the model architecture designed for prediction includes a main model and an auxiliary model. The main model is a pre-trained LSTM-Transformer hybrid network model to take into account both long short-term memory and global dependencies. The auxiliary model is a pre-trained temporal convolutional network model to extract local patterns.

[0030] In step S2, the LSTM-Transformer hybrid network model includes an LSTM layer and a Transformer encoder. After the feature tensor is input into the pre-trained LSTM-Transformer hybrid network model, it first enters the LSTM layer and then passes through the forget gate, input gate, cell state update, and output gate of the LSTM layer to generate the hidden state sequence. ,in The hidden state sequence is the output of the LSTM layer after processing the feature tensor. It is a set containing the hidden states at all time steps and is used as input to the Transformer encoder. , These represent the 1st time step, 2nd time step, ..., 1st time step of the LSTM layer, respectively. The single-step hidden state at each time step. (t=1,2,..., All are real number vectors with the same dimension as the number of hidden units H, storing key feature information for the corresponding time step.

[0031] The LSTM layer updates the LSTM cell state as follows: ; in Input features at time t, Let H be the hidden state at time t, and H be the number of hidden units. The cell state at time t. These are the trainable weight matrix and the bias term, respectively. It is the sigmoid activation function. This indicates element-wise multiplication.

[0032] The Transformer encoder receives a sequence of hidden states. First, a query vector, key vector, and value vector are generated through linear projection, represented as follows: ;in Both are projection matrices and .

[0033] Furthermore, in this embodiment, a self-attention mechanism is introduced into the Transformer encoder. Based on the query vector, key vector, and value vector, a sub-attention mechanism is used to calculate the global feature association, which is represented as follows: ;in For query vector dimensions.

[0034] In addition, this embodiment also constructs a position feedforward network in the Transformer encoder, represented as follows: ,in This is the input feature vector of the position feedforward network, i.e., the output of the self-attention mechanism layer in the Transformer encoder. This is the trainable weight matrix of the first layer of the location feedforward network, which serves to shift the input feature z from... 1. Mapping to higher hidden layer dimensions. This is the trainable weight matrix for the second layer of the position feedforward network, its function is to convert the output of the first layer... dimensional feature mapping back The dimension is consistent with the feature dimension of the input z. This is a trainable bias term for the first layer of the position feedforward network, used to adjust the linear transformation of the first layer. The baseline value is used to compensate for numerical offsets during the feature mapping process. This is a trainable bias term for the second layer of the position feedforward network, used to adjust the linear transformation of the second layer. The baseline value is used to ensure that the numerical range of the final output features is reasonable, and then the context code is output through a location feedforward network. .

[0035] In step S2, the temporal convolutional network model TCN includes causal convolutional layers, dilated convolutional layers, and activation function layers. Context encoding is input into the pre-trained temporal convolutional network model. First, it passes through the causal convolutional layers to ensure temporal consistency, then through the dilated convolutional layers to expand the receptive field, and finally through the ReLU activation function to output local temporal fusion features, represented as: .

[0036] In this embodiment, the Temporal Convolutional Network (TCN) model directly outputs ultra-short-term predictions of 5-30 minutes based on local temporal fusion features. Simultaneously, it extracts the features from the last time step in the local temporal fusion features and outputs short-term predictions of 1-24 hours through a fully connected layer. The expression for this is: ,in, This is the weight matrix. This is a bias term.

[0037] The final output is represented as: ;in This includes short-term forecasts of both load and renewable energy sources. This includes ultra-short-term forecasts that incorporate adjustable load baselines.

[0038] In this embodiment, a loss function is also constructed for the LSTM-Transformer hybrid network model and the temporal convolutional network model TCN. Specifically, a multi-task joint loss function is designed, including: 1) Based on Pinball Loss to handle uncertainty, the short-term prediction loss function is constructed as follows: ;in This represents the total number of short-term forecast periods, i.e., the total number of short-term time steps the model needs to predict. The quantile parameter controls the dimension by which PinballLoss captures prediction uncertainty; its value ranges from (0,1). These are the actual observed values ​​at time t, i.e., the true data of the short-term forecast target at time t. This represents the short-term predicted value for time period t, which is the short-term predicted result for time period t output by the LSTM-Transformer hybrid network model.

[0039] 2) The ultra-short-term prediction loss function constructed based on weighted MAE is as follows: , ;in This refers to the total number of adjustable loads covered by ultra-short-term forecasts, i.e., the total number of adjustable load devices participating in ultra-short-term regulation within the virtual power plant. This is the weighting coefficient for the i-th adjustable load, and its value is determined by the proportion of the maximum regulating capacity of this load to the total regulating capacity of all adjustable loads. This represents the actual observed value of the i-th adjustable load during time period t, i.e., the true operating data of this load during time period t. Let be the ultra-short-term forecast value of the i-th adjustable load in time period t, that is, the ultra-short-term forecast result of the load in time period t output by the sequential convolutional network model (TCN).

[0040] In this embodiment, the types of output prediction results include: 1) Adjustable load baseline: Its output form is , used to calculate and dynamically adjust constraint boundaries; 2) New energy power output: Its output form is , used to eliminate the construction of the objective function; 3) Total load demand: Its output form is It is used to achieve power balance constraints.

[0041] S3: Construct a multi-objective optimization model and model the constraints.

[0042] In step S3, the multi-objective optimization model includes maximizing the absorption of new energy sources, minimizing operating costs, minimizing carbon emissions, minimizing grid fluctuations, and maximizing user satisfaction as optimization objectives.

[0043] Based on the above optimization objective types, this embodiment first defines decision variables, including: a. with This represents the output of the controllable power source i during time period t, and its dimension is... Set its constraints as .

[0044] b, with Let J represent the charging and discharging power of energy storage j during time period t, respectively, with each dimension being 1. Set its constraints as .

[0045] c. with These represent the upward and downward adjustments of the adjustable load k during time period t, respectively, with each dimension being [missing information]. Set its constraints as .

[0046] d. with This represents the exchange power with the main network during time period t, and its dimension is... Set its constraints as .

[0047] e. with This represents the actual renewable energy power absorbed during time period t, and its dimension is... Set its constraints as .

[0048] in , , These represent the quantities of controllable power sources, energy storage, and adjustable loads, respectively. To optimize the total number of time periods, , These represent the minimum and maximum output power of power source i, respectively. This represents the maximum allowable adjustment amount for load k.

[0049] In step S3, the multi-objective optimization model is specifically represented as follows: ; ; ; ; ; in, To contribute to the predicted new energy sources, This refers to the actual amount of renewable energy power consumed. The electricity purchase price is t, and it is negative when the electricity is sold. and These represent the exchange power with the main network during time periods t and t-1, respectively. Let be the power generation cost coefficient of power source i. Let i be the output of power source i during time interval t. Let these represent the compensation costs for increasing and decreasing the load k, respectively. These represent the upward and downward adjustments of load k during time period t, respectively. Carbon intensity of grid electricity The carbon strength of power source i. All of these are comfort weights for load k.

[0050] In step S3, the constraints are modeled, including a power balance constraint model, a new energy consumption constraint model, an adjustable load regulation constraint model, an energy storage dynamic constraint model, and a controllable power source constraint model; specifically represented as follows: ; ; ; ; ; ; ; ; ; ; in, Let i be the output of power source i during time interval t. This refers to the actual amount of renewable energy power consumed. To contribute to the predicted new energy sources, Let be the charging and discharging power of energy storage j during time period t, respectively. The exchange power with the main network during time period t. This is the total load demand forecast. These represent the upward and downward adjustments to load k during time period t, respectively. These represent the maximum allowable upward and downward adjustments to load k, respectively. Let k be the operating efficiency coefficient for load k. The total number of time periods in the adjustment cycle. The time step for a single period. Let k be the baseline power. Let k be the total energy demand of load k. The temperature-to-power conversion factor, For the lowest comfortable temperature, The initial temperature. For the highest comfortable temperature, , Let J represent the state of charge of energy storage j at time points t and t-1, respectively. The charging efficiency of energy storage j, Let j be the discharge efficiency of the energy storage j. For the rated capacity of energy storage j, , These represent the minimum and maximum permissible states of charge (SOC) of energy storage j, respectively. , These represent the initial and final states of charge of energy storage j, respectively. This represents the maximum permissible ramp rate for power supply i. , Let be the start / stop state variables of power supply i at times t and t-1, respectively. Let i be the continuous operating time of power supply i. Let i be the minimum continuous operating time of power supply i.

[0051] Finally, the multi-objective optimization model is integrated based on the weighted summation method, and its expression is: ; ; in, The target weights are dynamically adjustable.

[0052] S4: Based on short-term and ultra-short-term prediction results, the multi-objective optimization model is solved through a pre-defined hierarchical architecture to obtain the optimal scheduling strategy and generate scheduling instructions.

[0053] In step S4, the specific method for solving the multi-objective optimization model using a preset hierarchical architecture is as follows: First, the hierarchical architecture includes a short-term scheduling layer and a real-time control layer. Then, based on the short-term prediction results, the short-term scheduling layer is used to solve the multi-objective optimization model using mixed-integer programming, outputting an hourly-level global scheduling strategy. Specifically, the short-term scheduling layer is used to solve the multi-objective optimization model using mixed-integer programming, and the continuous adjustment amount is discretized based on the adjustable load. Converting to integer values ​​simplifies the model, resulting in the following representation: ; in, The number of adjustable ranges for load k. The adjustment power corresponding to gear d is... These are binary decision variables.

[0054] An improved branch-and-bound method is used to optimize the short-term scheduling layer for 96 points throughout the day on an hourly time scale. This is achieved by modifying the quadratic term. and Piecewise linear approximation linearizes the objective function. Introducing auxiliary variables, represented as The acceleration strategies include: a. reusing the solution from the previous cycle as the initial solution for a warm start; b. using the Barrier method and concurrent MIP search to achieve parallel Gurobi solution.

[0055] Based on ultra-short-term prediction results, a real-time control layer is used to perform deep reinforcement learning online decision-making on a multi-objective optimization model, outputting a minute-level scheduling strategy; the deep reinforcement learning architecture includes: a. State space: ; b. Motion space: ; c. Reward function: ; in The power output of renewable energy sources for ultra-short-term forecasting, namely the minute-level power forecast of renewable energy sources (such as wind power and photovoltaics) based on the output of a temporal convolutional network (TCN) at the current decision moment, is a key feature in the state space reflecting the renewable energy supply capacity. The ultra-short-term baseline power of adjustable loads, i.e., the baseline operating power of each adjustable load predicted by the TCN without dispatch intervention, is the core data in the state space reflecting the basic load demand. The real-time control layer's scheduling and adjustment cycle, i.e., the time interval between a deep reinforcement learning decision and the output of a minute-level scheduling instruction, determines the real-time response frequency of the scheduling strategy. This refers to the ultra-short-term upward adjustment of the first adjustable load, which is the positive adjustment value output by the deep reinforcement learning decision used to adjust the operating power of the first adjustable load. For the first The ultra-short-term upward adjustment of adjustable load (of which) (This refers to the total number of adjustable loads within the virtual power plant), which is the positive adjustment value corresponding to the last adjustable load in the action space. The weighting of the reward for new energy consumption. As the weight of the sub-reward for operating costs, To constrain the weight of penalty items for violations, Sub-rewards for the consumption of new energy. As a sub-reward for operating costs, To constrain violations, penalties are imposed for violations of system constraints (such as exceeding the upper limit of load regulation or exceeding the limit of energy storage SOC).

[0056] The formula for setting up a reward for new energy consumption is as follows: ,in The actual amount consumed; penalties for breach of contract are set as follows: .

[0057] Finally, based on the hourly global scheduling strategy and the minute-level scheduling strategy, the optimal scheduling strategy is obtained.

[0058] In this embodiment, a near-end strategy is used to optimize the PPO for minute-level scheduling policy updates, and the formula is as follows: ;in Indicates will Limited to Within the range.

[0059] S5: Distributes scheduling instructions to distributed resources within the Virtual Power Plant (VPP), enabling precise execution and conversion into equipment actions through the local controller.

[0060] In step S5, a closed-loop feedback is also constructed to ensure the safety of dynamic adjustment, specifically including: 1) The communication protocol is designed in layers, including an adjustable load layer, an energy storage system layer, and a controllable power supply layer. After the adjustable load command is converted, the equipment is dynamically adjusted according to the command based on an adaptive PID control algorithm, as shown below: , in Define the error and , Indicates that the gain is dynamically adjusted and , This is the rated power of the energy storage. , , All of these are basic PID parameters obtained through equipment calibration.

[0061] 2) Configure the load balancing strategy, expressed as follows: , Let k be the priority weight of device k. The device is in an available state. This represents the actual adjustment amount performed by the k-th adjustable load during time period t, i.e., the final power adjustment value that this load needs to perform after load balancing. The model output instruction adjustment amount for the k-th adjustable load during time period t is the initial scheduling instruction generated after solving the multi-objective optimization model for that load. The priority weight for the k-th adjustable load is a preset weighting coefficient based on factors such as equipment importance and adjustment capacity, used to determine the allocation ratio of the adjustment amount. Let be the priority weight of the j-th adjustable load in the set of available devices Ω. Let $\frac{j}{j}$ be the availability status variable for the j-th adjustable load, taking a value of 1 or 0. 1 indicates that the equipment is operating normally and can execute adjustment commands; 0 indicates that the equipment is faulty, offline, or otherwise unavailable and does not participate in the allocation of adjustment amounts. This is the set of available equipment for adjustable loads.

[0062] 3) Introduce a safety boundary crossing protection mechanism, including: a. Real-time power mutation constraint: ;in This represents the measured power value of the k-th device during time period t, which is the actual operating power of the device collected in real time by sensors. It is used to monitor whether power changes exceed the safe range. Let be the measured power value of the k-th device during time period t-1 (i.e., the previous measurement time period), and be... The power change between adjacent time periods is calculated together. Let be the maximum permissible power change rate for the k-th device, a safety threshold determined by the device's hardware performance, limiting the rate of power fluctuation. The time step is the time interval between two adjacent power measurement periods, i.e., the time interval between two power measurements by the acquisition device, used to convert the power change rate into the actual power change.

[0063] b. ,in This refers to the real-time temperature of the equipment.

[0064] c. SOC Security Corridor: .

[0065] In this embodiment, prediction errors and optimization deviations are dynamically corrected through real-time feedback data to achieve closed-loop control, ensuring that the system maintains optimal operation under uncertainties such as new energy fluctuations and sudden load changes. The steps include: 1) Input multi-source data and provide feedback at the execution layer. Introducing disturbance factors such as changes in outdoor temperature and light intensity, Data fusion processing based on exponential smoothing fusion is expressed as: ,in As a smoothing factor, This is the predicted value for the previous period; where The data set represents the multi-dimensional measured data from the execution layer at time t, including the measured power of the k-th device, the measured state of charge of the j-th energy storage device, and the measured temperature of the k-th device. This data forms the foundation for reflecting the real-time operating status of the system. This represents the measured state of charge (SOC) of the j-th energy storage device, which is the ratio of its current remaining charge to its rated capacity, collected in real-time by sensors. This represents the change in outdoor temperature during time period t, that is, the difference between the outdoor temperature in time period t and the outdoor temperature in the previous time period. Let be the change in light intensity during time period t, that is, the difference between the light intensity during time period t and the light intensity during the previous time period. The data at time t is the composite data after exponential smoothing.

[0066] 2) Modeling of new energy prediction errors, expressed as: ; The correction formula is expressed as follows ; Where W is the sliding window width and k is the correction gain. The rate of change of error and , The actual renewable energy power absorbed during time period t (such as the actual grid-connected power of wind power and photovoltaics) serves as the baseline true value for calculating the renewable energy forecast error. Comparison with the forecast value reflects the magnitude of the error. The standard deviation of the new energy forecast error for time period t is calculated by taking the square root of the mean of the squared historical errors within the sliding window width W. It is used to quantify the uncertainty of new energy forecasts (the larger the standard deviation, the more significant the forecast fluctuation). for Forecasted output of new energy sources after time-adjusted period for The original forecast value of new energy power output for the time period.

[0067] 3) The online update of the optimization model parameters is based on adaptive adjustment of the target weights, and the formula is as follows: ;in Let be the weighting coefficient of the new energy consumption target at time t. The weighting coefficients for the renewable energy consumption target at the initial time (t=0) are... This is the weight adjustment factor (learning rate), used to control the magnitude of weight updates. This represents the total number of new energy devices within the virtual power plant. Let be the actual output (or actual power consumption) of the i-th renewable energy device during time period t, used to calculate the actual consumption / predicted output ratio of the device.

[0068] The formula for updating the policy gradient based on real-time data is as follows: ;in For policy network parameters The gradient of the objective function represents the objective function. (e.g., maximizing cumulative rewards) on parameters The set of partial derivatives is the core basis for updating the policy network through gradient ascent. In Deep Reinforcement Learning (DRL), the set of trainable parameters for the policy network includes the weight matrices and bias terms of each layer, which determine the decision-making logic of the policy network to output actions based on the state. The action output by the policy network at time t, i.e., the minute-level scheduling decision of the real-time control layer, is the output result of the policy network. Let be the system state vector at time t, which includes energy storage SOC, ultra-short-term predicted output of new energy sources, and baseline load power, etc., and is the input data for the policy network. Let be the adjusted reward value at time t. This is the state value function, used to evaluate the state at time t. The long-term expected reward.

[0069] 4) The rolling optimization process includes: a. Status update: acquire real-time data and call the prediction model to refresh the prediction; b. Optimization solution: if it is within the feasible region of DRL, output the action directly; if it is close to the constraint boundary, trigger the MIP solver to perform correction; c. Instruction generation: discretize the continuous actions into device instructions. 5) Feedback correction based on Kalman filtering is expressed as: Equations of state:

[0070] Observation equation:

[0071] in , , Here is the system state transition matrix. To control the input matrix, Let be the control input vector at time t, i.e., the set of dispatch instructions for the virtual power plant. The process noise at time t, Let be the observation vector at time t. For the observation matrix, The observation noise at time t, For generating power from new energy sources (predicted or actual), used to track dynamic changes in new energy power output. It provides real-time equipment temperature (such as the operating temperature of load and energy storage) to monitor equipment operation safety (e.g., power needs to be limited if the temperature is too high).

[0072] Introducing constraint violation metric, denoted as ,in The constraint violation degree of the k-th device at time t. Summation of time index, At the current target time for calculating the degree of constraint violation, The sliding window width is the size of the time window used to calculate the degree of constraint violation. The actual adjustment amount of the kth device. This represents the maximum allowable adjustment amount for the k-th device, which is the upper limit of the device adjustment constraint.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A virtual power plant load regulation method based on deep learning scheduling strategy, characterized in that, include: Acquire multi-source heterogeneous data and preprocess the multi-source heterogeneous data to obtain feature tensors; The feature tensor is input into a pre-trained LSTM-Transformer hybrid network model, and the output is a context code; the context code is input into a pre-trained temporal convolutional network model, and local temporal fusion features are extracted to output ultra-short-term prediction results. Extract the last time step feature from the local temporal fusion features, and output the short-term prediction result through a fully connected layer; A multi-objective optimization model is constructed, and the constraints are modeled. Based on short-term and ultra-short-term prediction results, the multi-objective optimization model is solved through a pre-defined hierarchical architecture that includes a short-term scheduling layer and a real-time control layer to obtain the optimal scheduling strategy and generate scheduling instructions. The scheduling instructions are sent to the distributed resources within the virtual power plant, executed, and transformed into equipment actions.

2. The virtual power plant load regulation method based on deep learning scheduling strategy as described in claim 1, characterized in that, The multi-source heterogeneous data includes historical load data, renewable energy output data, meteorological data, electricity price and market data, and equipment status data; The preprocessing includes missing value imputation, outlier detection, dimensionality elimination, and feature alignment; wherein, missing value imputation uses linear interpolation or the nearest neighbor mean method, outlier detection uses interquartile range method, dimensionality elimination uses data standardization, and feature alignment aligns all data according to a unified timestamp.

3. The virtual power plant load regulation method based on deep learning scheduling strategy as described in claim 1, characterized in that, The LSTM-Transformer hybrid network model includes an LSTM layer and a Transformer encoder. The feature tensor is passed through the forget gate, input gate, cell state update, and output gate of the LSTM layer to generate a hidden state sequence. The Transformer encoder receives the hidden state sequence, first generates a query vector, key vector, and value vector through linear projection, then calculates global feature association through a sub-attention mechanism, and finally outputs the context code through a position feedforward network.

4. The virtual power plant load regulation method based on deep learning scheduling strategy as described in claim 1, characterized in that, The temporal convolutional network model includes causal convolutional layers, dilated convolutional layers, and activation function layers. Context encoding is input into the pre-trained temporal convolutional network model. First, it passes through the causal convolutional layer to ensure the temporal order is reasonable. Then, it passes through the dilated convolutional layer to expand the receptive field. Finally, it outputs local temporal fusion features through the ReLU activation function.

5. The virtual power plant load regulation method based on deep learning scheduling strategy as described in claim 1, characterized in that, The multi-objective optimization model includes the following objectives: maximizing renewable energy absorption, minimizing operating costs, minimizing carbon emissions, minimizing grid fluctuations, and maximizing user satisfaction; specifically, it is expressed as follows: ； ； ； ； ； in, To contribute to the predicted new energy sources, This refers to the actual amount of renewable energy power consumed. The electricity purchase price is t, and it is negative when the electricity is sold. and These represent the exchange power with the main network during time periods t and t-1, respectively. Let be the power generation cost coefficient of power source i. Let i be the output of power source i during time interval t. Let these represent the compensation costs for increasing and decreasing the load k, respectively. These represent the upward and downward adjustments of load k during time period t, respectively. Carbon intensity of grid electricity The carbon strength of power source i. All of these are comfort weights for load k.

6. The virtual power plant load regulation method based on deep learning scheduling strategy as described in claim 1, characterized in that, The modeling of constraints includes power balance constraint model, new energy consumption constraint model, adjustable load regulation constraint model, energy storage dynamic constraint model, and controllable power source constraint model; specifically represented as follows: ； ； ； ； ； ； ； ； ； ； in, Let i be the output of power source i during time interval t. This refers to the actual amount of renewable energy power consumed. To contribute to the predicted new energy sources, Let be the charging and discharging power of energy storage j during time period t, respectively. The exchange power with the main network during time period t. This is the total load demand forecast. These represent the upward and downward adjustments to load k during time period t, respectively. These represent the maximum allowable upward and downward adjustments to load k, respectively. Let k be the operating efficiency coefficient for load k. The total number of time periods in the adjustment cycle. The time step for a single period. Let k be the baseline power. Let k be the total energy demand of load k. The temperature-to-power conversion factor, For the lowest comfortable temperature, The initial temperature. For the highest comfortable temperature, , Let J represent the state of charge of energy storage j at time points t and t-1, respectively. The charging efficiency of energy storage j, Let j be the discharge efficiency of the energy storage j. For the rated capacity of energy storage j, , These represent the minimum and maximum permissible states of charge (SOC) of energy storage j, respectively. , These represent the initial and final states of charge of energy storage j, respectively. This represents the maximum permissible ramp rate for power supply i. , Let be the start / stop state variables of power supply i at times t and t-1, respectively. Let i be the continuous operating time of power supply i. Let i be the minimum continuous operating time of power supply i.

7. The virtual power plant load regulation method based on deep learning scheduling strategy as described in claim 1, characterized in that, The specific method for solving the multi-objective optimization model using a pre-defined hierarchical architecture that includes a short-term scheduling layer and a real-time control layer is as follows: First, the hierarchical architecture includes a short-term scheduling layer and a real-time control layer. Then, based on the short-term prediction results, the short-term scheduling layer is used to solve the multi-objective optimization model using mixed integer programming, and an hourly global scheduling strategy is output. Based on the ultra-short-term prediction results, a real-time control layer is used to perform deep reinforcement learning online decision-making on the multi-objective optimization model, outputting a minute-level scheduling strategy; finally, based on the hour-level global scheduling strategy and the minute-level scheduling strategy, the optimal scheduling strategy is obtained.

8. A virtual power plant load regulation system based on deep learning scheduling strategy, characterized in that, include: The data acquisition module is configured to: acquire multi-source heterogeneous data and preprocess the multi-source heterogeneous data to obtain feature tensors; The prediction module is configured to: input feature tensors into a pre-trained LSTM-Transformer hybrid network model and output context encoding; input the context encoding into a pre-trained temporal convolutional network model, extract local temporal fusion features, and output ultra-short-term prediction results; extract the last time step features from the local temporal fusion features and output short-term prediction results through a fully connected layer. The model building and solving module is configured to: build a multi-objective optimization model and model the constraints; based on short-term and ultra-short-term prediction results, solve the multi-objective optimization model through a preset hierarchical architecture to obtain the optimal scheduling strategy and generate scheduling instructions; The execution module is configured to: issue the scheduling instructions to distributed resources within the virtual power plant, execute them, and convert them into equipment actions.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the virtual power plant load regulation method based on deep learning scheduling strategy as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the virtual power plant load regulation method based on deep learning scheduling strategy as described in any one of claims 1-7.

Images