Virtual power plant edge intelligent cooperative control method and device

Abstract

The invention discloses a virtual power plant edge intelligent cooperative control method and device, and belongs to the field of virtual power plant control. Comprising the steps that multi-source data of distributed energy and loads in a virtual power plant are collected and preprocessed, and a time sequence characteristic matrix is obtained; inputting the time sequence characteristic matrix into a lightweight deep learning model deployed on an edge AI chip for reasoning operation, and obtaining a load prediction result, a power generation prediction result and an anomaly detection result in a future time period; constructing a rolling optimization problem of a limited prediction time domain on the edge side and solving the rolling optimization problem to obtain an energy storage charging and discharging power instruction, a flexible load regulation instruction and a power grid interaction power instruction at the current moment; and the 5G-RedCap communication module issues a power instruction and an adjustment instruction to field execution equipment, so that real-time cooperative control of the virtual power plant is realized. The real-time performance, the intelligent level and the operation reliability of the virtual power plant are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of virtual power plant control, and particularly relates to a virtual power plant edge intelligent collaborative control method and device. Background Technology

[0002] In response to energy conservation and emission reduction, distributed photovoltaic (PV) power, energy storage systems, electric vehicle charging stations, and flexible loads are rapidly penetrating various industrial parks. The power system is undergoing a profound transformation from the traditional "centralized large-scale generating units + passive loads" architecture to a "distributed renewable energy + active loads" model. Virtual power plant (VPS) technology, as a core means of integrating massive distributed resources, incorporates dispersed power sources, energy storage, and loads into a unified scheduling and trading framework through communication networks, achieving functions such as peak shaving, demand response, and ancillary services. Meanwhile, 5G RedCap lightweight communication technology, with its low latency, high reliability, and wide coverage, provides a new solution for massive edge terminal access in IoT scenarios. Deep learning algorithms demonstrate powerful nonlinear feature extraction capabilities in processing high-dimensional, multi-source heterogeneous data, and are widely used in load forecasting, output forecasting, and anomaly detection. Some existing technologies have combined edge computing architecture with VPS control, attempting to deploy lightweight models close to the data source to alleviate cloud pressure, and have initially explored AI-driven local decision-making capabilities, laying the technological foundation for the intelligent evolution of VPS.

[0003] However, existing virtual power plant control systems still face multiple bottlenecks: First, traditional 4G public networks or dedicated lines are prone to link congestion in large-scale access scenarios, making it difficult to compress end-to-end latency to sub-second levels, which cannot match the second-level fluctuations in photovoltaic output and the rapid charging and discharging response requirements of energy storage devices; second, most solutions adopt a centralized architecture that is "heavy on the cloud and light on the edge," requiring all data to be uploaded to the cloud for unified processing, which not only increases the cloud's computing power burden but also causes the site to lose its autonomy when the network fluctuates or is interrupted, leading to the risk of local loss of control; third, existing systems mostly use fixed parameter models trained offline. The existing scheduling strategies (such as ARIMA or simple machine learning models) lack online incremental learning and parameter adaptation mechanisms. When the seasons change, the load structure evolves, or the equipment ages, the prediction accuracy and control performance rapidly decline. In addition, most existing scheduling strategies focus on single economic goals and rarely incorporate multi-dimensional indicators such as carbon emission factors, green certificate benefits, and equipment operation safety into a unified optimization framework, making it difficult to achieve a dynamic balance between low carbon emissions, economy, and safety. Finally, the system lacks robustness and lacks effective offline autonomy and smooth switching mechanisms under abnormal operating conditions such as communication interruptions, which limits the overall reliability and continuous operation capability. Summary of the Invention

[0004] The purpose of this invention is to overcome the problems of high communication latency, slow response due to centralized architecture, inability of deep learning models to adapt online, and insufficient consideration of carbon emissions and multi-objective optimization in existing virtual power plant control systems. This invention proposes a deep learning-based 5G-RedCap virtual power plant edge intelligent collaborative control method and device to achieve the following objectives:

[0005] Achieve millisecond-level energy consumption prediction and control strategy generation at the edge;

[0006] Low-latency, high-reliability interaction between the virtual power plant cloud platform and edge devices is achieved through the 5G-RedCap network;

[0007] By integrating online incremental learning with multi-edge node parameter fusion, the model maintains its adaptability to changes in seasons, holidays, and device status.

[0008] By comprehensively considering economic efficiency, carbon emissions, and operational safety within the end-edge-cloud collaborative framework, the overall benefits of virtual power plants can be improved.

[0009] To achieve the above objectives, the present invention provides a virtual power plant edge intelligent collaborative control method and apparatus. The virtual power plant edge intelligent collaborative control method includes:

[0010] The system acquires multi-source operation data of distributed energy and load in a virtual power plant, and performs filtering, normalization and time synchronization processing on the multi-source operation data to generate a time series feature matrix.

[0011] A lightweight deep learning model is constructed, and inference operations are performed on the time series feature matrix based on the lightweight deep learning model to obtain the load prediction power, distributed power output prediction power, carbon emission prediction and anomaly detection score for multiple future time steps.

[0012] Based on the predicted load power, the predicted output power of distributed power sources, and the predicted carbon emissions, combined with real-time electricity price information, energy storage status, and flexible load priority, a rolling optimization problem with a finite prediction time domain is constructed at the edge node. The optimal control sequence for energy storage charging and discharging power, flexible load start and stop commands, and grid interaction power is obtained by solving the rolling optimization problem. The first time step control command in the optimal control sequence is then executed.

[0013] The prediction results, execution command records, and operating status are reported to the cloud virtual power plant platform through the 5G-RedCap communication module. The global optimization strategy issued by the cloud virtual power plant platform is received, and the local control of the edge nodes is corrected or constrained according to the global optimization strategy.

[0014] When the divergence of the input data distribution relative to the historical distribution exceeds a preset threshold, the online incremental learning of the lightweight deep learning model is triggered. The weights of some network layers of the model are fine-tuned to generate incremental parameters. The incremental parameters are uploaded to the cloud virtual power plant platform to participate in federated aggregation, and the updated global model parameters are received from the cloud broadcast.

[0015] When communication is interrupted or the cloud is unavailable, the edge device automatically switches to autonomous control mode, performs security boundary control based on local prediction results and policy library, and gradually returns to cloud-edge collaborative control mode after communication is restored.

[0016] Preferably, the process of generating the time series feature matrix includes:

[0017] Collect photovoltaic inverter operation data, energy storage BMS data, electric vehicle charging pile status data, building load meter data, ambient temperature and humidity, light intensity, wind speed and carbon emission coefficient;

[0018] The collected data undergoes outlier removal, missing value imputation, data normalization, and multi-channel time synchronization to generate a normalized time series feature matrix for model inference.

[0019] Preferably, the lightweight deep learning model includes an improved CNN-LSTM hybrid neural network model;

[0020] The process of constructing a lightweight deep learning model and performing inference operations on the time series feature matrix based on the lightweight deep learning model includes:

[0021] Based on the convolutional feature extraction layer, a depthwise separable convolutional structure is used to perform channel-wise spatial convolution and point convolution channel mixing on the input time series feature matrix;

[0022] After depthwise separable convolution, an SE channel attention module is introduced to perform global average pooling, fully connected mapping and sigmoid activation operations on the channel dimension of the convolution output to obtain the weight coefficients of each channel and recalibrate the features of each channel.

[0023] Based on the temporal modeling layer, an LSTM network is used to model the recalibrated feature sequence. A temporal attention mechanism is introduced on the hidden state sequence output by the LSTM to calculate the attention weights at each time step and to perform weighted aggregation on the hidden state sequence to generate a context vector.

[0024] The context vector is received through a fully connected prediction layer, which outputs the predicted load power, predicted carbon emissions, and anomaly detection score for future time steps.

[0025] Preferably, the process of constructing a rolling optimization problem in the finite prediction time domain at the edge nodes includes:

[0026] The objective function is the weighted sum of the grid purchase cost, carbon emission cost and power smoothing penalty term in the future prediction time domain. The constraints are set as follows: power balance constraint, upper limit constraint of energy storage charging and discharging power, energy storage SOC dynamic equation constraint, SOC upper and lower limit constraint and feeder capacity constraint.

[0027] The rolling optimization method of model predictive control is adopted. In each control cycle, the rolling optimization problem is solved with the current state as the initial condition to obtain the optimal control sequence for multiple future time steps. Only the control instruction of the first time step is executed, and the rolling optimization problem is reconstructed and solved in the next control cycle.

[0028] The independent variables of the objective function include energy storage charging power, energy storage discharging power, flexible load start-stop state variables, and grid power purchase.

[0029] Preferably, the 5G-RedCap communication module establishes bidirectional communication with the cloud-based virtual power plant platform through a dedicated 5G-RedCap network or a virtual private network, supports MQTT, IEC 104, and Modbus-TCP protocols, performs QoS classification and network slicing mapping according to service type, and achieves real-time communication with end-to-cloud round-trip latency of less than 50ms.

[0030] Preferably, the triggering conditions for the online incremental learning include:

[0031] Based on calendar and environmental information, the operating days are divided into different scenario categories, and a baseline distribution of statistical characteristics for each scenario category is established.

[0032] The latest observed features are collected within a sliding window to form an empirical distribution, and the Kullback-Leibler divergence between the empirical distribution and the baseline distribution is calculated.

[0033] Model fine-tuning is triggered when the Kullback-Leibler divergence exceeds a preset threshold.

[0034] Monitor the equipment health index. When the equipment health exceeds the normal range and the predicted residual variance exceeds the threshold, trigger an incremental model update.

[0035] Preferably, the online incremental learning process includes:

[0036] Only the LSTM layer weight parameters are updated, using the Adam optimizer with a learning rate of 1e-4, and the number of update steps is limited to no more than 100.

[0037] Every 24 hours, the edge nodes upload the incremental parameters obtained from the fine-tuning to the cloud virtual power plant platform. The cloud performs federated aggregation of the incremental parameters uploaded by multiple edge nodes to obtain the updated global model parameters and broadcast them to all edge nodes.

[0038] Preferably, the autonomous control mode is as follows:

[0039] When the 5G-RedCap communication module detects a communication interruption or cloud unavailability, the edge device activates the security boundary control policy in the local policy library and continues to perform energy storage charging and discharging and load regulation control based on the prediction results of the local CNN-LSTM model to maintain power supply security.

[0040] After communication is restored, the edge device gradually receives the global optimization strategy from the cloud-based virtual power plant platform, progressively adjusts its local control parameters, and returns to the cloud-edge collaborative control mode.

[0041] The present invention also provides a virtual power plant edge intelligent collaborative control device, comprising:

[0042] The data acquisition module is used to acquire multi-source operational data of distributed energy resources and loads in the virtual power plant;

[0043] The data processing module, connected to the data acquisition module, is used to filter, normalize, and synchronize the multi-source running data to generate a time series feature matrix.

[0044] The inference and prediction module, connected to the data processing module, is used to construct a lightweight deep learning model and perform inference operations on the time series feature matrix based on the lightweight deep learning model to obtain the predicted load power, predicted distributed power output, predicted carbon emissions, and anomaly detection score for multiple future time steps.

[0045] The instruction execution module, connected to the inference and prediction module, is used to construct a rolling optimization problem in the finite prediction time domain at the edge node based on the predicted load power, the predicted output power of distributed power sources, the predicted carbon emissions, and combined with real-time electricity price information, energy storage status, and flexible load priority. The rolling optimization problem is solved to obtain the optimal control sequence of energy storage charging and discharging power, flexible load start and stop commands, and grid interaction power, and the first time step control command in the optimal control sequence is executed.

[0046] The communication interaction module, connected to the instruction execution module, is used to report the prediction results, execution instruction records and operating status to the cloud virtual power plant platform through the 5G-RedCap communication module, receive the global optimization strategy issued by the cloud virtual power plant platform, and correct or constrain the local control of the edge nodes according to the global optimization strategy.

[0047] The model update module, connected to the communication interaction module, is used to trigger online incremental learning of the lightweight deep learning model when the divergence of the input data distribution relative to the historical distribution exceeds a preset threshold. This module fine-tunes the weights of some network layers in the model to generate incremental parameters, uploads the incremental parameters to the cloud virtual power plant platform to participate in federated aggregation, and receives the updated global model parameters broadcast from the cloud.

[0048] The mode switching module, connected to the model update module, is used to automatically switch the edge device to the autonomous control mode when communication is interrupted or the cloud is unavailable. It performs security boundary control based on local prediction results and policy library, and gradually returns to the cloud-edge collaborative control mode after communication is restored.

[0049] Compared with the prior art, the present invention has the following advantages and technical effects:

[0050] Compared with existing technologies, this invention achieves millisecond-level edge intelligent prediction and real-time control response. Through 5G-RedCap communication, it reduces end-to-cloud round-trip latency to less than 50ms, significantly improving system real-time performance. Adopting an "edge-cloud" collaborative architecture, it deploys deep learning inference and local optimization at the edge, effectively reducing the computing load on the cloud and automatically switching to local autonomous control mode when communication is interrupted, enhancing system robustness and reliability. The introduction of online incremental learning and federated parameter update mechanisms enables the model to adaptively fine-tune according to seasonal, holiday, and equipment status changes, improving prediction accuracy by more than 40% over long-term operation. It integrates economic efficiency, carbon emissions, and power smoothing multi-dimensional indicators into the optimization objective, achieving a dynamic balance of comprehensive benefits. The overall solution supports modular deployment and large-scale expansion, reducing peak energy consumption by 12%–15% in practical applications in zero-carbon parks, with a system stability rate of 99.6%. Attached Figure Description

[0051] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0052] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0053] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0054] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0055] Example 1

[0056] like Figure 1 As shown, this embodiment provides a virtual power plant edge intelligent collaborative control method, including:

[0057] The system acquires multi-source operation data of distributed energy and load in a virtual power plant, performs filtering, normalization and time synchronization processing on the multi-source operation data, and generates a time series feature matrix.

[0058] A lightweight deep learning model is constructed, and inference operations are performed on the time series feature matrix based on the lightweight deep learning model to obtain the load prediction power, distributed power generation prediction power, carbon emission prediction and anomaly detection score for multiple future time steps.

[0059] Based on the predicted load power, the predicted output power of distributed power sources, and the predicted carbon emissions, combined with real-time electricity price information, energy storage status, and flexible load priority, a rolling optimization problem with a finite prediction time domain is constructed at the edge node. The optimal control sequence for energy storage charging and discharging power, flexible load start and stop commands, and grid interaction power is obtained by solving the rolling optimization problem. The first time step control command in the optimal control sequence is then executed.

[0060] The 5G-RedCap communication module reports the prediction results, execution command records and operating status to the cloud virtual power plant platform, receives the global optimization strategy issued by the cloud virtual power plant platform, and corrects or constrains the local control of the edge nodes according to the global optimization strategy.

[0061] When the divergence of the input data distribution relative to the historical distribution exceeds a preset threshold, the online incremental learning of the lightweight deep learning model is triggered. The weights of some network layers of the model are fine-tuned to generate incremental parameters. The incremental parameters are uploaded to the cloud virtual power plant platform to participate in federated aggregation and receive the updated global model parameters broadcast from the cloud.

[0062] When communication is interrupted or the cloud is unavailable, the edge device automatically switches to autonomous control mode, performs security boundary control based on local prediction results and policy library, and gradually returns to cloud-edge collaborative control mode after communication is restored.

[0063] Furthermore, the process of generating the time series feature matrix includes:

[0064] Collect photovoltaic inverter operation data, energy storage BMS data, electric vehicle charging pile status data, building load meter data, ambient temperature and humidity, light intensity, wind speed and carbon emission coefficient;

[0065] The collected data undergoes outlier removal, missing value imputation, data normalization, and multi-channel time synchronization to generate a normalized time series feature matrix for model inference.

[0066] Furthermore, this embodiment performs multi-channel synchronous data acquisition on various distributed energy sources and loads within the park, and performs data preprocessing to complete noise reduction, alignment, and feature construction, generating a time-series feature matrix for model inference.

[0067] Furthermore, lightweight deep learning models include improved CNN-LSTM hybrid neural network models;

[0068] The process of constructing a lightweight deep learning model and performing inference operations on the time series feature matrix based on the lightweight deep learning model includes:

[0069] Based on the convolutional feature extraction layer, a depthwise separable convolutional structure is used to perform channel-wise spatial convolution and point convolution channel mixing on the input time series feature matrix;

[0070] After depthwise separable convolution, an SE channel attention module is introduced to perform global average pooling, fully connected mapping and sigmoid activation operations on the channel dimension of the convolution output to obtain the weight coefficients of each channel and recalibrate the features of each channel.

[0071] Based on the temporal modeling layer, an LSTM network is used to model the recalibrated feature sequence. A temporal attention mechanism is introduced on the hidden state sequence output by the LSTM to calculate the attention weights at each time step and to perform weighted aggregation on the hidden state sequence to generate a context vector.

[0072] The fully connected prediction layer receives the context vector and outputs the predicted load power, carbon emission prediction, and anomaly detection score for future time steps.

[0073] Furthermore, this embodiment, based on the conventional CNN+LSTM framework, makes the following module replacements and improvements to the convolutional feature extraction layer and the temporal modeling layer: First, in the convolutional feature extraction layer, depthwise separable convolution is used instead of standard convolution, and a channel attention module (Squeeze-and-Excitation, SE module) is connected in series after the convolution output to improve feature representation capability on computationally limited edge chips; Second, in the temporal modeling layer, a temporal attention mechanism is introduced on the hidden state sequence output by LSTM to form an attention-enhanced LSTM (Att-LSTM) structure, which adaptively weights and aggregates the hidden states at different time steps, thereby highlighting key time segments that are more sensitive to load changes and abnormal operating conditions.

[0074] The improved CNN-LSTM hybrid neural network described in this embodiment is used to simultaneously capture the spatial and temporal features of energy consumption data. The network structure consists of five main layers, as follows:

[0075] ① Input layer;

[0076] Input data type: Time series energy consumption matrix;

[0077] ;

[0078] Where T is the time step (e.g., 96 points / day at a 15-minute granularity), and n is the number of features (e.g., voltage, current, temperature, irradiance, carbon emission coefficient, etc.).

[0079] Example of input dimensions: T=96, n=12.

[0080] ② The convolutional feature extraction layer is an improved CNNBlock;

[0081] The convolutional feature extraction layer employs a depthwise separable convolutional structure, meaning it first performs channel-wise spatial convolution on each feature channel, and then... Point convolution performs channel blending, with a kernel size of [value missing]. It has 32 output channels and uses ReLU activation. Compared to ordinary convolution, this structure significantly reduces the number of parameters and multiplication / accumulation operations, making it more suitable for deployment on embedded AI chips.

[0082] This embodiment introduces a SE channel attention module after depthwise separable convolution. This module performs "global average pooling → two fully connected layers → sigmoid activation" on the channel dimensions of the convolution output to obtain the weight coefficients of each channel. And recalibrate the characteristics of each channel:

[0083]

[0084] With the above structure, this embodiment can automatically enhance the focus on key feature channels such as drastic load changes and abnormal fluctuations under limited computing power, and weaken redundant or noise features, thereby improving the overall prediction performance.

[0085] ③ Temporal modeling layer: Attention-enhanced LSTMBlock;

[0086] The temporal modeling layer uses a single-layer or two-layer LSTM network to model the convolutional feature sequence, with 128 hidden units. Assume the time step formed after rearrangement of the convolutional output is... Feature sequences The hidden state sequence generated by LSTM is denoted as Its basic update equation is the same as that of the standard LSTM.

[0087] To further enhance the model's attention to key time segments, this embodiment introduces a time attention mechanism based on the LSTM output sequence:

[0088] ;

[0089] Where, vector ,matrix and bias For trainable parameters, For the first Attention weights at each time step This is the weighted context vector. Compared to directly using the hidden state of the last time step, this embodiment uses an attention-weighted aggregation method, which allows the model to adaptively select the time segment that contributes most to the prediction, thereby improving the accuracy and robustness of long-sequence load and carbon emission prediction.

[0090] ④ Fully connected prediction layer (DenseLayer);

[0091] Output dimension: 3

[0092] Output metrics:

[0093] Power forecast for the next 15 minutes;

[0094] Corresponding carbon emission forecast;

[0095] Anomaly detection score;

[0096] Activation function: Linear (used for regression);

[0097] Output equation:

[0098] ;

[0099] ⑤ Policy generation layer;

[0100] Specifically, the control strategy generation layer implements a multi-input multi-output nonlinear control function. Its inputs include: the current state of charge of the energy storage. CNN-LSTM predicts future short-term load power vector Predicting carbon emission vectors Real-time electricity prices and their deviation from benchmark electricity prices and anomaly detection score etc.; the output is a reference value for energy storage charging and discharging power. Flexible load comprehensive adjustment coefficient and abnormal linkage control flags .

[0101] In its implementation, the control policy generation layer uses a two-layer fully connected network. Perform parameterization:

[0102] ;

[0103] Wherein, the input vector is:

[0104] ;

[0105] For ReLU or Leaky-ReLU activation functions, For learnable parameters; output vector After linear transformation and saturation constraint, they are mapped as follows:

[0106] ;

[0107] in, Indicates in Saturation function within the range, This indicates that the output will be restricted to... interval, For indicator functions, This is a preset alarm threshold.

[0108] The control strategy generation layer in this embodiment employs a multi-parameter nonlinear control function, rather than a simple single threshold comparison. This control function takes the prediction results of CNN-LSTM, energy storage SOC, real-time electricity price changes, and anomaly detection scores as inputs. It uses a two-layer fully connected network to approximate the analytical control law for the local rolling optimization problem, achieving joint decision-making on energy storage charging and discharging power and flexible load start-up and shutdown signals. By introducing multi-dimensional state variables and prediction variables as independent variables of the control function, the control strategy generation layer in this embodiment can more precisely balance economy (electricity price), safety (SOC boundary), and anomaly risk while ensuring real-time performance. This significantly improves the intelligence and technical effectiveness of the control strategy, distinguishing it from existing control methods that rely solely on simple thresholds or rule bases.

[0109] The improved CNN-LSTM hybrid neural network model in this embodiment does not simply adopt the existing standard CNN+LSTM stacked structure. Instead, it introduces structural replacements and improvements such as "depthiable separable convolution + channel attention module" and "attention-enhanced LSTM (Att-LSTM)" in the convolutional feature extraction layer and temporal modeling layer, respectively.

[0110] Depthwise separable convolutions are used instead of traditional convolutions in the convolutional layers, and a Squeeze-and-Excitation channel attention module is introduced to enhance the adaptive selection capability of key feature channels while keeping the number of parameters and computational cost low.

[0111] In the time series layer, a time attention mechanism is added after the LSTM output sequence to perform weighted aggregation of the hidden states at different times, highlighting the time segments that are sensitive to load changes, thereby improving prediction accuracy.

[0112] Through the above design, the strategy layer of this embodiment no longer relies on a single threshold judgment, but constructs a complex nonlinear control function with multidimensional prediction quantities and state quantities as independent variables, realizes joint optimization decision-making for energy storage charging and discharging and flexible load regulation, and makes the control function approximate the optimal control law of the local rolling optimization problem through offline training and online fine-tuning, taking into account both real-time performance and control quality.

[0113] This embodiment calls a CNN-LSTM network to predict active power, reactive power, load level and carbon emissions for multiple future time steps, and to score the current operating conditions to determine equipment failure, communication failure or energy consumption failure.

[0114] Furthermore, the process of constructing a rolling optimization problem in the finite prediction time domain at the edge nodes includes:

[0115] The objective function is the weighted sum of the grid purchase cost, carbon emission cost and power smoothing penalty term in the future prediction time domain. The constraints are set as follows: power balance constraint, upper limit constraint of energy storage charging and discharging power, energy storage SOC dynamic equation constraint, SOC upper and lower limit constraint and feeder capacity constraint.

[0116] The rolling optimization method of model predictive control is adopted. In each control cycle, the rolling optimization problem is solved with the current state as the initial condition to obtain the optimal control sequence for multiple future time steps. Only the control instruction of the first time step is executed, and the rolling optimization problem is reconstructed and solved in the next control cycle.

[0117] The independent variables of the objective function include energy storage charging power, energy storage discharging power, flexible load start-stop state variables, and grid power purchase.

[0118] Furthermore, this embodiment uses the future output of the CNN-LSTM model... Load forecast power at each time step Distributed power generation output prediction power and carbon emission forecasts Combined with real-time electricity prices Information such as the upper and lower limits of energy storage SOC and the priority of flexible loads is used to construct the following local rolling optimization problem in the finite prediction time domain at the edge nodes:

[0119]

[0120] in, , These are the energy storage charging and discharging power, respectively. For the first Start-stop status of an adjustable load. To purchase power from the grid, , These are the carbon emission cost weighting and power smoothing weighting coefficients.

[0121] The optimization problem described above satisfies the following constraints:

[0122]

[0123] ;

[0124] in, For the battery pack in the first The state of charge at each time step , These are the charging and discharging efficiencies, respectively. For rated capacity, For time step, For the first The flexible load has adjustable power. This is the upper limit of the feeder capacity.

[0125] The optimization problem described above is a linearly constrained optimization problem with a quadratic penalty term. In this embodiment, it is solved using the rolling optimization method of Model Predictive Control (MPC): within each control cycle, the problem is solved using the current state as the initial condition to obtain the future... The optimal control sequence for each time step is determined by executing only the energy storage charging and discharging power, flexible load start-stop command, and electric vehicle charging power setpoint for the first time step. In the next control cycle, the updated measurements are used to reconstruct and solve the optimization problem.

[0126] The solution process is implemented using an embedded optimization solver, and the solution time is controlled within tens of milliseconds per cycle, thereby ensuring the real-time performance of the edge device under a 1Hz control cycle. The control commands obtained from the solution are finally sent by the execution control unit (50) to the energy storage inverter, load controller and charging pile to realize peak shaving and valley filling, demand response and carbon emission optimization control.

[0127] Furthermore, the 5G-RedCap communication module establishes bidirectional communication with the cloud-based virtual power plant platform through a dedicated 5G-RedCap network or a virtual private network, supports MQTT, IEC 104, and Modbus-TCP protocols, performs QoS classification and network slicing mapping according to service type, and achieves real-time communication with end-to-cloud round-trip latency of less than 50ms.

[0128] Furthermore, in this embodiment, the prediction results, execution command records, and operating status are reported to the cloud virtual power plant platform through the 5G-RedCap communication module. The cloud platform performs higher-level optimizations based on global information (such as cross-park power sharing, power trading, and carbon asset settlement), and then distributes the global optimization strategy to each edge node to correct or constrain its local control.

[0129] Furthermore, the triggering conditions for online incremental learning include:

[0130] Based on calendar and environmental information, the operating days are divided into different scenario categories, and a baseline distribution of statistical characteristics for each scenario category is established.

[0131] The latest observed features are collected within a sliding window to form an empirical distribution, and the Kullback-Leibler divergence between the empirical distribution and the baseline distribution is calculated.

[0132] Model fine-tuning is triggered when the Kullback-Leibler divergence exceeds a preset threshold.

[0133] Monitor the equipment health index. When the equipment health exceeds the normal range and the predicted residual variance exceeds the threshold, trigger an incremental model update.

[0134] Furthermore, the model parameters and deployment optimizations in this embodiment are shown in Table 1.

[0135] Table 1

[0136] Module Number of parameters Key hyperparameters computational complexity Reasoning Time (Jetson Nano) CNN layer <![CDATA[3.8×10 4 ]]> Convolution kernel = 3 × 1, channels = 32 O(T×n×k) 8ms LSTM layer <![CDATA[1.1×10 5 ]]> Hidden layer nodes = 128, time steps = 48 O(T×128²) 15ms Dense layer 384 Activation = Linear O(1) 1ms overall <![CDATA[≈1.5×10 5 ]]> - - <25ms

[0137] The edge deployment optimization described in this embodiment uses the TensorRT accelerated inference framework, combined with half-precision (FP16) computation and dynamic batch size control, to achieve millisecond-level response times in 5G-RedCap networks.

[0138] Furthermore, the edge device in this embodiment possesses an online incremental learning function to adapt to different seasons, holidays, or changes in device status. The mechanism is as follows:

[0139] Data caching and feature drift detection (correlation with seasons / holidays / equipment status): To enable the deep learning model to adapt to different seasons and energy usage patterns, this embodiment first divides the operating day into scenarios based on calendar and environmental information:

[0140] The year is divided into three categories based on external meteorological conditions: winter, summer, and spring / autumn transition season, with the daily average temperature range serving as the basis for the division.

[0141] Based on energy consumption habits, operating days are further divided into three categories: "weekdays, weekends, and statutory holidays".

[0142] This allows for the creation of various scenario combinations, such as "summer weekdays" and "winter holidays." During the initial operation phase, the system statistically analyzes the baseline distribution of characteristics such as load power, indoor and outdoor temperature, and occupancy rate for each scenario. .

[0143] During online operation, the edge device automatically determines the current scene category based on the current date and ambient temperature. and in length of The latest observation features are collected within a sliding window to form an empirical distribution. By calculating the Kullback–Leibler divergence between the two:

[0144] ;

[0145] When the divergence exceeds a preset threshold This indicates that the load pattern under the current "season + calendar" scenario has changed significantly, such as the combination of high summer temperatures and new load access or concentrated electricity consumption during holidays. At this time, it triggers a local model fine-tuning for this scenario.

[0146] In addition, this embodiment also introduces equipment status-related indicators, such as the estimated equivalent charge-discharge efficiency of the energy storage system, the number of charge-discharge cycles, the frequency of use within the SOC operating range, and the frequency of fault alarms of key equipment, to statistically obtain the equipment health index. .when If the error exceeds the normal range (e.g., a significant decrease or a significant increase in alarm frequency) and causes the variance of the prediction residual to exceed the threshold within the current sliding window, then the error is considered to be outside the normal range. In addition, incremental updates to the model will be triggered to adapt to the impact of equipment aging or replacement on the dynamic characteristics of the system.

[0147] Specifically, this embodiment divides the operating days into different scenarios according to "spring / summer / autumn / winter" and "weekday / weekend / holiday", and statistically analyzes the reference data distribution for each scenario. In the edge node, the current scenario category is determined based on the current calendar information and ambient temperature, and only the sliding window data corresponding to the scenario is compared with the corresponding reference distribution for divergence. When the divergence exceeds the threshold, a local fine-tuning is triggered for the scenario, thereby achieving adaptation under seasonal and holiday modes.

[0148] Meanwhile, this embodiment introduces equipment health status indicators (such as estimated energy storage charging and discharging efficiency, fault alarm frequency, etc.). When the equipment status changes significantly, causing abnormal prediction residual statistical characteristics, the model will also be updated.

[0149] Through the above design, the self-learning and feature drift detection mechanism is no longer an abstract distribution comparison, but establishes a clear correspondence with "energy consumption patterns in different seasons and holidays" and "changes in equipment status", thereby ensuring that the model update is targeted and has scene perception capabilities.

[0150] Furthermore, the local fine-tuning involved in this implementation includes:

[0151] Only update the LSTM layer weights Wh, UhW_h, U_hWh, Uh;

[0152] Learning rate: 1e-4;

[0153] Optimizer: Adam;

[0154] Limit the number of update steps to ≤100 to reduce computational burden.

[0155] Furthermore, the cloud synchronization involved in this implementation includes:

[0156] The edge side uploads a new parameter ΔW every 24 hours;

[0157] The cloud integrates parameters from multiple edge nodes to generate a globally optimal model.

[0158] The cloud broadcasts the new model to all edge nodes, enabling federated updating.

[0159] Furthermore, the process of implementing online incremental learning includes:

[0160] Only the LSTM layer weight parameters are updated, using the Adam optimizer with a learning rate of 1e-4, and the number of update steps is limited to no more than 100.

[0161] Every 24 hours, the edge nodes upload the incremental parameters obtained from the fine-tuning to the cloud virtual power plant platform. The cloud performs federated aggregation of the incremental parameters uploaded by multiple edge nodes to obtain the updated global model parameters and broadcast them to all edge nodes.

[0162] Furthermore, when the divergence of the input data distribution relative to the historical distribution exceeds a preset threshold, the edge device automatically triggers online fine-tuning of some network layers (such as LSTM layers) to form incremental parameters; the cloud platform periodically collects the incremental parameters uploaded by each edge node, performs federated aggregation, obtains the updated global model, and broadcasts it to all edge nodes.

[0163] Furthermore, the autonomous control model is as follows:

[0164] When the 5G-RedCap communication module detects a communication interruption or cloud unavailability, the edge device activates the security boundary control policy in the local policy library and continues to perform energy storage charging and discharging and load regulation control based on the prediction results of the local CNN-LSTM model to maintain power supply security.

[0165] After communication is restored, the edge device gradually receives the global optimization strategy from the cloud-based virtual power plant platform, progressively adjusts its local control parameters, and returns to the cloud-edge collaborative control mode.

[0166] Furthermore, in this embodiment, when communication is interrupted or the cloud is unavailable, the edge device automatically switches to autonomous control mode and continues to implement security boundary control based on local prediction and policy library. After communication is restored, it gradually returns to cloud-edge collaborative control mode.

[0167] The method in this embodiment can perform edge intelligent prediction and achieve millisecond-level response: Specifically, by deploying a pruned and quantized CNN-LSTM network on the edge side and using inference acceleration frameworks such as TensorRT, the latency of a single inference is controlled within tens of milliseconds. Combined with a control cycle of 1Hz to 10Hz, fine-grained real-time control can be achieved.

[0168] The 5G-RedCap communication module in this embodiment has the advantages of low latency and high reliability communication. With the help of 5G-RedCap optimization for IoT scenarios, the system can achieve real-time communication with end-to-cloud round-trip latency of less than 50ms while ensuring controllable terminal cost and power consumption, providing a foundation for rapid coordination and scheduling of virtual power plants.

[0169] The method of this embodiment, which involves a three-layer collaboration between the end, edge, and cloud, improves overall efficiency and robustness. This embodiment deploys the data acquisition and execution functions at the end layer, the deep learning inference and local optimization at the edge layer, and the global optimization and model training at the cloud layer, realizing an architecture of "nearby decision-making + global coordination". This effectively alleviates the computing power pressure on the cloud and maintains the autonomous control capability of the edge side in the event of communication failure.

[0170] The method in this embodiment supports online incremental learning and federated model updates. Edge devices can make local fine-tuning based on changes in data distribution and participate in cloud-based federated learning by uploading incremental parameters, enabling the model to continuously absorb information from new scenarios and operating conditions, thereby improving prediction accuracy and control performance over long-term operation.

[0171] The method in this embodiment takes into account economy, low carbon emissions and safety. It introduces multi-dimensional indicators such as peak-valley time-of-use electricity price, carbon emission factor, equipment operation constraints and user comfort into the control objective function. Through multi-objective weighted or hierarchical optimization, it can achieve a dynamic balance between saving electricity costs, reducing carbon emissions and ensuring energy supply security.

[0172] The method in this embodiment is easy to deploy and expand in engineering. The device in this embodiment adopts a modular design in hardware and a containerized deployment and pluggable model framework in software. It can flexibly connect to different brands of equipment and multiple communication protocols, which facilitates rapid large-scale promotion in new and existing parks.

[0173] Example 2

[0174] Based on the same inventive concept, this embodiment also provides a virtual power plant edge intelligent collaborative control device, including:

[0175] The data acquisition module is used to acquire multi-source operational data of distributed energy resources and loads in the virtual power plant;

[0176] The data processing module, connected to the data acquisition module, is used to filter, normalize, and synchronize time-series running data from multiple sources to generate a time series feature matrix.

[0177] The inference and prediction module, connected to the data processing module, is used to build a lightweight deep learning model. Based on the lightweight deep learning model, it performs inference operations on the time series feature matrix to obtain the predicted load power, the predicted output power of distributed power sources, the predicted carbon emissions, and the anomaly detection score for multiple future time steps.

[0178] The instruction execution module, connected to the inference and prediction module, is used to construct a rolling optimization problem in the finite prediction time domain at the edge node based on the predicted load power, the predicted output power of distributed power sources, the predicted carbon emissions, and combined with real-time electricity price information, energy storage status, and flexible load priority. The rolling optimization problem is solved to obtain the optimal control sequence of energy storage charging and discharging power, flexible load start and stop instructions, and grid interaction power, and the first time step control instruction in the optimal control sequence is executed.

[0179] The communication interaction module, connected to the instruction execution module, is used to report the prediction results, execution instruction records and operating status to the cloud virtual power plant platform through the 5G-RedCap communication module, receive the global optimization strategy issued by the cloud virtual power plant platform, and correct or constrain the local control of the edge nodes according to the global optimization strategy.

[0180] The model update module, connected to the communication interaction module, is used to trigger online incremental learning of the lightweight deep learning model when the divergence of the input data distribution relative to the historical distribution exceeds a preset threshold. It fine-tunes the weights of some network layers of the model to generate incremental parameters, uploads the incremental parameters to the cloud virtual power plant platform to participate in federated aggregation, and receives the updated global model parameters broadcast from the cloud.

[0181] The mode switching module, connected to the model update module, is used to automatically switch the edge device to autonomous control mode when communication is interrupted or the cloud is unavailable. It performs security boundary control based on local prediction results and policy library, and gradually returns to cloud-edge collaborative control mode after communication is restored.

[0182] Furthermore, the data acquisition module in this embodiment is used to collect photovoltaic inverter operation data, energy storage BMS data, electric vehicle charging pile status, building load meter data, as well as multi-source data such as ambient temperature and humidity, light intensity, wind speed, and carbon emission coefficient.

[0183] Furthermore, the inference prediction module in this embodiment is configured with an embedded AI chip with GPU / NPU, and deploys a CNN-LSTM hybrid deep neural network model that has been lightly pruned and quantized to complete multi-timescale load and output prediction, carbon emission prediction, and abnormal operating condition identification, and further generate local control strategies; specifically, by calling the CNN-LSTM hybrid deep neural network model, it predicts active power, reactive power, load level, and carbon emissions for multiple future time steps, and performs anomaly scoring on the current operating condition to determine equipment failure, communication anomaly, or energy consumption anomaly.

[0184] This embodiment deploys a CNN-LSTM hybrid deep neural network model on an embedded AI chip, enabling millisecond-level prediction.

[0185] Furthermore, the communication interaction module in this embodiment is used to conduct bidirectional communication with the virtual power plant cloud platform through a 5G-RedCap dedicated or virtual private network, supports protocols such as MQTT, IEC104, and Modbus-TCP, and can perform QoS classification and network slice mapping according to service type;

[0186] This embodiment utilizes a 5G-RedCap communication module for low-latency communication to achieve bidirectional synchronization between the park's energy equipment and the virtual power plant's main control center.

[0187] Furthermore, the mode switching module in this embodiment includes an execution control unit and a local storage unit;

[0188] The execution control unit is used to electrically connect with field devices such as energy storage inverters, charging piles, air conditioning and lighting systems, and interruptible / transferable loads through PWM modules, relay arrays, and analog / digital output interfaces, and is used to perform control actions such as peak shaving and valley filling, demand response, and reserve capacity release.

[0189] The local storage unit is used to store historical sampling data, model parameters, incremental update weights, control logs and security audit records, and supports ring overlay and data encryption;

[0190] This embodiment achieves a combination of autonomy and global control, as well as three-layer collaboration between the edge, cloud, and end-user, through data acquisition, edge inference, and cloud optimization. Furthermore, the edge device can fine-tune model parameters online to adapt to changes in environment and load.

[0191] The edge device's operation cycle is: one sampling prediction and control command update per second (1Hz cycle), and the model inference time is <30ms, ensuring millisecond-level real-time response.

[0192] Furthermore, this embodiment also includes a power management module for providing multiple regulated outputs, surge protection, and backup power support to maintain the operation of critical device functions in the event of an external power failure.

[0193] Furthermore, this embodiment also includes a housing and an interface module, which provide structural protection for the device with a protection level of not less than IP54, and provide physical interfaces such as RS485, CAN, Ethernet, DI / DO, and SIM / USIM card slots.

[0194] Furthermore, in this embodiment, the device is installed at the energy control node of the virtual power plant park, and collaborative operation is achieved through the following connection:

[0195] The data acquisition module is electrically connected to equipment such as photovoltaic inverters, energy storage BMS, electric vehicle charging pile controllers, and electricity meters in the park.

[0196] The acquired signals are sent to the inference and prediction module via the data processing module. The inference and prediction module communicates bidirectionally with the virtual power plant control center via the 5G-RedCap communication module.

[0197] The control signal is fed back to the field execution equipment through the mode switching module to realize closed-loop control of energy flow;

[0198] All modules are powered by a unified power management module;

[0199] The data is stored in local storage units and periodically uploaded to the cloud.

[0200] Twenty units of this device were deployed in a zero-carbon smart park for the group control of energy storage cabinets and charging piles. Results showed that the prediction error (MAE) was 2.3%, a 41% improvement over the traditional model; peak energy consumption was reduced by 12%–15%; response latency was shortened to <50ms; and the system's stable operation rate reached 99.6%.

[0201] The device in this embodiment reduces the mean absolute error (MAE) of prediction accuracy to 2.3%, which is better than the traditional LSTM model (3.9%). The response latency is ≤10ms from the end to the edge and ≤40ms from the edge to the cloud through 5GRedCap communication. The energy consumption is reduced by about 12% to 15% after peak shaving and valley filling optimization. Furthermore, it can still maintain local self-control in the event of data drift and communication interruption.

[0202] As an additional implementation, this embodiment also provides a virtual power plant edge intelligent collaborative control system, which includes at least:

[0203] Several edge intelligent collaborative control devices as described above;

[0204] Cloud-based virtual power plant control platform;

[0205] Several field-controllable devices, including photovoltaic inverters, energy storage devices, charging piles, building automation systems, industrial load controllers, etc.;

[0206] All edge devices are connected to the cloud-based virtual power plant control platform via the 5G-RedCap network, forming a multi-node hierarchical collaborative scheduling structure.

[0207] The device system in this embodiment can be deployed in zero-carbon smart parks, industrial parks, data center parks, and urban integrated energy systems to realize multi-timescale prediction, optimized scheduling, and adaptive intelligent control of distributed energy.

[0208] The virtual power plant edge intelligent collaborative control device provided in this embodiment has all the advantages of the virtual power plant edge intelligent collaborative control method provided in Embodiment 1.

[0209] Example 3

[0210] This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in Embodiment 1.

[0211] Example 4

[0212] This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0213] Example 5

[0214] This embodiment also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0215] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A virtual power plant edge intelligent collaborative control method, characterized in that, include: The system acquires multi-source operation data of distributed energy and load in a virtual power plant, and performs filtering, normalization and time synchronization processing on the multi-source operation data to generate a time series feature matrix. A lightweight deep learning model is constructed, and inference operations are performed on the time series feature matrix based on the lightweight deep learning model to obtain the load prediction power, distributed power output prediction power, carbon emission prediction and anomaly detection score for multiple future time steps. Based on the predicted load power, the predicted output power of distributed power sources, and the predicted carbon emissions, combined with real-time electricity price information, energy storage status, and flexible load priority, a rolling optimization problem with a finite prediction time domain is constructed at the edge node. The optimal control sequence for energy storage charging and discharging power, flexible load start and stop commands, and grid interaction power is obtained by solving the rolling optimization problem. The first time step control command in the optimal control sequence is then executed. The prediction results, execution command records, and operating status are reported to the cloud virtual power plant platform through the 5G-RedCap communication module. The global optimization strategy issued by the cloud virtual power plant platform is received, and the local control of the edge nodes is corrected or constrained according to the global optimization strategy. When the divergence of the input data distribution relative to the historical distribution exceeds a preset threshold, the online incremental learning of the lightweight deep learning model is triggered. The weights of some network layers of the model are fine-tuned to generate incremental parameters. The incremental parameters are uploaded to the cloud virtual power plant platform to participate in federated aggregation, and the updated global model parameters are received from the cloud broadcast. When communication is interrupted or the cloud is unavailable, the edge device automatically switches to autonomous control mode, performs security boundary control based on local prediction results and policy library, and gradually returns to cloud-edge collaborative control mode after communication is restored.

2. The method according to claim 1, characterized in that, The process of generating a time series feature matrix includes: Collect photovoltaic inverter operation data, energy storage BMS data, electric vehicle charging pile status data, building load meter data, ambient temperature and humidity, light intensity, wind speed and carbon emission coefficient; The collected data undergoes outlier removal, missing value imputation, data normalization, and multi-channel time synchronization to generate a normalized time series feature matrix for model inference.

3. The method according to claim 1, characterized in that, The lightweight deep learning model includes an improved CNN-LSTM hybrid neural network model; The process of constructing a lightweight deep learning model and performing inference operations on the time series feature matrix based on the lightweight deep learning model includes: Based on the convolutional feature extraction layer, a depthwise separable convolutional structure is used to perform channel-wise spatial convolution and point convolution channel mixing on the input time series feature matrix; After depthwise separable convolution, an SE channel attention module is introduced to perform global average pooling, fully connected mapping and sigmoid activation operations on the channel dimension of the convolution output to obtain the weight coefficients of each channel and recalibrate the features of each channel. Based on the temporal modeling layer, an LSTM network is used to model the recalibrated feature sequence. A temporal attention mechanism is introduced on the hidden state sequence output by the LSTM to calculate the attention weights at each time step and to perform weighted aggregation on the hidden state sequence to generate a context vector. The context vector is received through a fully connected prediction layer, which outputs the predicted load power, predicted carbon emissions, and anomaly detection score for future time steps.

4. The method according to claim 1, characterized in that, The process of constructing a rolling optimization problem with a finite prediction time domain at edge nodes includes: The objective function is the weighted sum of the grid purchase cost, carbon emission cost and power smoothing penalty term in the future prediction time domain. The constraints are set as follows: power balance constraint, upper limit constraint of energy storage charging and discharging power, energy storage SOC dynamic equation constraint, SOC upper and lower limit constraint and feeder capacity constraint. The rolling optimization method of model predictive control is adopted. In each control cycle, the rolling optimization problem is solved with the current state as the initial condition to obtain the optimal control sequence for multiple future time steps. Only the control instruction of the first time step is executed, and the rolling optimization problem is reconstructed and solved in the next control cycle. The independent variables of the objective function include energy storage charging power, energy storage discharging power, flexible load start-stop state variables, and grid power purchase.

5. The method according to claim 1, characterized in that, The 5G-RedCap communication module establishes bidirectional communication with the cloud-based virtual power plant platform through a dedicated 5G-RedCap network or a virtual private network. It supports MQTT, IEC 104, and Modbus-TCP protocols, performs QoS classification and network slicing mapping according to service type, and achieves real-time communication with end-to-cloud round-trip latency of less than 50ms.

6. The method according to claim 1, characterized in that, The triggering conditions for the online incremental learning include: Based on calendar and environmental information, the operating days are divided into different scenario categories, and a baseline distribution of statistical characteristics for each scenario category is established. The latest observed features are collected within a sliding window to form an empirical distribution, and the Kullback-Leibler divergence between the empirical distribution and the baseline distribution is calculated. Model fine-tuning is triggered when the Kullback-Leibler divergence exceeds a preset threshold. Monitor the equipment health index. When the equipment health exceeds the normal range and the predicted residual variance exceeds the threshold, trigger an incremental model update.

7. The method according to claim 1, characterized in that, The execution process of the online incremental learning includes: Only the LSTM layer weight parameters are updated, using the Adam optimizer with a learning rate of 1e-4, and the number of update steps is limited to no more than 100. Every 24 hours, the edge nodes upload the incremental parameters obtained from the fine-tuning to the cloud virtual power plant platform. The cloud performs federated aggregation of the incremental parameters uploaded by multiple edge nodes to obtain the updated global model parameters and broadcast them to all edge nodes.

8. The method according to claim 1, characterized in that, The autonomous control mode is as follows: When the 5G-RedCap communication module detects a communication interruption or cloud unavailability, the edge device activates the security boundary control policy in the local policy library and continues to perform energy storage charging and discharging and load regulation control based on the prediction results of the local CNN-LSTM model to maintain power supply security. After communication is restored, the edge device gradually receives the global optimization strategy from the cloud-based virtual power plant platform, progressively adjusts its local control parameters, and returns to the cloud-edge collaborative control mode.

9. A virtual power plant edge intelligent collaborative control device, characterized in that, include: The data acquisition module is used to acquire multi-source operational data of distributed energy resources and loads in the virtual power plant; The data processing module, connected to the data acquisition module, is used to filter, normalize, and synchronize the multi-source running data to generate a time series feature matrix. The inference and prediction module, connected to the data processing module, is used to construct a lightweight deep learning model and perform inference operations on the time series feature matrix based on the lightweight deep learning model to obtain the predicted load power, predicted distributed power output, predicted carbon emissions, and anomaly detection score for multiple future time steps. The instruction execution module, connected to the inference and prediction module, is used to construct a rolling optimization problem in the finite prediction time domain at the edge node based on the predicted load power, the predicted output power of distributed power sources, the predicted carbon emissions, and combined with real-time electricity price information, energy storage status, and flexible load priority. The rolling optimization problem is solved to obtain the optimal control sequence of energy storage charging and discharging power, flexible load start and stop commands, and grid interaction power, and the first time step control command in the optimal control sequence is executed. The communication interaction module, connected to the instruction execution module, is used to report the prediction results, execution instruction records and operating status to the cloud virtual power plant platform through the 5G-RedCap communication module, receive the global optimization strategy issued by the cloud virtual power plant platform, and correct or constrain the local control of the edge nodes according to the global optimization strategy. The model update module, connected to the communication interaction module, is used to trigger online incremental learning of the lightweight deep learning model when the divergence of the input data distribution relative to the historical distribution exceeds a preset threshold. This module fine-tunes the weights of some network layers in the model to generate incremental parameters, uploads the incremental parameters to the cloud virtual power plant platform to participate in federated aggregation, and receives the updated global model parameters broadcast from the cloud. The mode switching module, connected to the model update module, is used to automatically switch the edge device to the autonomous control mode when communication is interrupted or the cloud is unavailable. It performs security boundary control based on local prediction results and policy library, and gradually returns to the cloud-edge collaborative control mode after communication is restored.

Images