Neural network-based off-grid micro-grid system load prediction method and system

Abstract

The invention belongs to the technical field of load prediction, and discloses an off-grid micro-grid system load prediction method and system based on a neural network. The method comprises the following steps: constructing an equipment fault diagnosis model, an energy scheduling model and a load prediction model by using a neural network algorithm; carrying out data acquisition and preprocessing to obtain preprocessed real-time area monitoring data; performing equipment fault diagnosis by using the equipment fault diagnosis model; if the real-time equipment fault diagnosis result is that a fault exists, entering an energy scheduling step, otherwise, entering a load prediction step; performing energy scheduling by using the energy scheduling model, executing the obtained real-time energy scheduling scheme, and returning to the data acquisition step; and performing load prediction by using the load prediction model. According to the method, the problems of data island, insufficient model generalization ability, limited fault diagnosis precision and efficiency, insufficient energy scheduling strategy optimization and poor load prediction precision and robustness in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of load prediction, and particularly relates to a load prediction method and system for an off-grid micro-grid system based on a neural network. BACKGROUND

[0002] A micro-grid system is a small-scale power generation and distribution system composed of distributed power sources, energy storage devices, energy conversion devices, loads (or loads), monitoring and protection devices, and other micro-grid equipment. Among them, the off-grid micro-grid system can realize autonomous operation without external grid support. With the wide application of renewable energy, the off-grid micro-grid system plays an important role in remote areas and specific application scenarios. However, the existing technology still has many defects, including: 1) Data island and poor model generalization ability: traditional neural network models are scattered in each regional base station, lacking effective data sharing mechanism; models independently trained by each base station are difficult to utilize global data, resulting in poor model generalization ability and weak adaptability to new scenarios; for example, a fault diagnosis model trained by a base station according to its own historical data may not be able to accurately identify when facing other base station-specific equipment fault types or load conditions, resulting in misdiagnosis or missed diagnosis; 2) Limited fault diagnosis precision and efficiency: existing fault diagnosis methods mostly rely on threshold judgment or shallow machine learning models, which are difficult to effectively handle complex nonlinear relationships and high-dimensional data; the ability to extract early weak fault features is insufficient, resulting in low diagnosis precision and prone to false positives and false negatives; in addition, traditional methods often require manual analysis of large amounts of data, resulting in low diagnosis efficiency and difficulty in meeting real-time requirements; 3) Insufficient optimization of energy dispatching strategy: traditional energy dispatching strategies are mostly rule-based empirical models or simple optimization algorithms, which are difficult to consider the complex characteristics and randomness of multi-source heterogeneous energy and the dynamic constraints of off-grid micro-grid system operation; in scenarios with high renewable energy penetration, existing methods are difficult to achieve optimal balance between economy and reliability, resulting in low energy utilization and high operating costs; 4) Poor load prediction precision and robustness: existing load prediction methods mostly use single models, which are difficult to effectively handle multi-source heterogeneous data and nonlinear relationships; the adaptability to complex weather conditions and changes in user behavior is poor, resulting in low prediction precision, especially when facing sudden load changes, the prediction result deviation is large. SUMMARY

[0003] In order to solve the problems of data island and poor model generalization ability, limited fault diagnosis precision and efficiency, insufficient optimization of energy dispatching strategy, and poor load prediction precision and robustness in the existing technology, the present application aims to provide a load prediction method and system for an off-grid micro-grid system based on a neural network.

[0004] The technical scheme adopted by the present application is: A neural network-based load prediction method for an off-grid micro-grid system, comprising the following steps: Using a neural network algorithm, a device fault diagnosis model, an energy scheduling model, and a load prediction model for the off-grid micro-grid system are constructed. Data collection is performed, and the collected real-time regional monitoring data of the off-grid micro-grid system in each monitoring area is preprocessed to obtain preprocessed real-time regional monitoring data. According to the preprocessed real-time regional monitoring data, a device fault diagnosis model is used to perform device fault diagnosis, and real-time device fault diagnosis results are obtained. If the real-time device fault diagnosis result is a fault, proceed to the energy scheduling step, otherwise integrate all preprocessed real-time regional monitoring data to obtain real-time system monitoring data and proceed to the load prediction step. According to the real-time device fault diagnosis result and the preprocessed real-time regional monitoring data, an energy scheduling model is used to perform energy scheduling, and the obtained real-time energy scheduling scheme is executed, and the data collection step is returned. According to the real-time system monitoring data of the off-grid micro-grid system, a load prediction model is used to perform load prediction, and real-time load prediction results are obtained.

[0005] Further, using a neural network algorithm, a device fault diagnosis model, an energy scheduling model, and a load prediction model for the off-grid micro-grid system are constructed, comprising the following steps: Build a cloud-edge collaborative architecture for the off-grid micro-grid system; the cloud-edge collaborative architecture includes a cloud data center, regional base stations in several monitoring areas, and several edge computing gateways in each monitoring area. Based on the cloud data center, an initial device fault diagnosis model and an initial load prediction model are constructed using a deep learning algorithm, and an initial energy scheduling model is constructed using a swarm intelligence optimization algorithm. According to the continuous learning mechanism, set the experience replay pool and the comprehensive loss function of the initial device fault diagnosis model, the initial load prediction model and the initial energy scheduling model. Based on the cloud-edge collaborative architecture, according to the federated learning mechanism, training optimization is performed to obtain the final device fault diagnosis model, the final load prediction model and the final energy scheduling model.

[0006] Further, the device fault diagnosis model is constructed based on the CNN-LSTM algorithm.

[0007] Further, the load prediction model is constructed based on the GCN-RF-DBN-SVM algorithm.

[0008] Further, the energy scheduling model is constructed based on the FSSA algorithm.

[0009] Further, data collection is performed, and the collected real-time regional monitoring data of the off-grid micro-grid system in each monitoring area is preprocessed to obtain preprocessed real-time regional monitoring data, including the following steps: Collecting heterogeneous real-time equipment monitoring data of different data sources in each monitoring area of the off-grid micro-grid system; According to the preset data format, the data mapping and conversion are performed on the plurality of heterogeneous real-time equipment monitoring data to obtain a plurality of homogeneous real-time equipment monitoring data; The plurality of homogeneous real-time equipment monitoring data is sequentially subjected to data cleaning, Gaussian denoising and normalization processing to obtain a plurality of standard real-time equipment monitoring data; The plurality of standard real-time equipment monitoring data is subjected to data dimension reduction to obtain a plurality of reduced real-time equipment monitoring data, and the plurality of reduced real-time equipment monitoring data is integrated to obtain the preprocessed real-time regional monitoring data.

[0010] Further, according to the preprocessed real-time regional monitoring data, the equipment fault diagnosis model is used to perform equipment fault diagnosis to obtain real-time equipment fault diagnosis results, including the following steps: The preprocessed real-time regional monitoring data including the plurality of reduced real-time equipment monitoring data is converted into a real-time regional monitoring data matrix; Using the equipment fault diagnosis model, real-time matrix features of the real-time regional monitoring data matrix are extracted; According to the real-time matrix features, equipment fault diagnosis is performed to obtain real-time equipment fault diagnosis results; If the real-time equipment fault diagnosis result is that there is a fault, the energy scheduling step is entered, otherwise all the preprocessed real-time regional monitoring data is integrated to obtain real-time system monitoring data, and the load prediction step is entered.

[0011] Further, according to the real-time equipment fault diagnosis result and the preprocessed real-time regional monitoring data, the energy scheduling model is used to perform energy scheduling, and the obtained real-time energy scheduling scheme is executed, and the data collection step is returned, including the following steps: Using the energy scheduling model, the real-time equipment fault diagnosis result is analyzed to obtain real-time influence factors, and according to the real-time influence factors, a real-time optimization target is set; According to the real-time optimization target, the preset fitness function of the energy scheduling model is updated to obtain a real-time fitness function; According to the real-time equipment fault diagnosis result and the preprocessed real-time regional monitoring data, a real-time energy scheduling scheme format and a real-time constraint condition are set; According to the real-time energy scheduling scheme format and the real-time constraint condition, initial solution generation is carried out to obtain a plurality of initial solutions; the initial solution corresponds to an initial real-time energy scheduling scheme; According to the real-time fitness function and the real-time constraint condition, the plurality of initial solutions are iteratively updated, and an optimal solution with an optimal real-time fitness value is output; The individual vector of the optimal solution is subjected to solution vector analysis to obtain an optimal real-time energy scheduling scheme; The optimal real-time energy scheduling scheme is published to the off-grid micro-grid system, the optimal real-time energy scheduling scheme is executed, and the data collection step is returned.

[0012] Further, according to the real-time system monitoring data of the off-grid micro-grid system, a load prediction model is used to perform load prediction to obtain a real-time load prediction result, including the following steps: The load prediction model is used to extract real-time graph structure features of the real-time system monitoring data; A plurality of real-time key features of the real-time graph structure features are extracted; According to the plurality of real-time key features, load prediction diagnosis is performed to obtain a real-time load prediction probability distribution; According to the plurality of real-time key features, error compensation prediction is performed to obtain a real-time error compensation value; The real-time load prediction probability distribution and the corresponding real-time error compensation value are added to obtain a real-time load prediction result of the off-grid micro-grid system; Updating is performed to obtain an updated equipment fault diagnosis model, an updated energy scheduling model and an updated load prediction model.

[0013] A neural network-based off-grid micro-grid system load prediction system is used to implement an off-grid micro-grid system load prediction method, and the system includes a cloud data center, a plurality of regional base stations, a plurality of edge computing gateways, a plurality of monitoring data collection devices and a plurality of micro-grid devices.

[0014] The beneficial effects of the present application are: This invention provides a load forecasting method and system for off-grid microgrid systems based on neural networks. It employs a coalition learning mechanism, allowing base stations in different regions to collaboratively train by exchanging model parameters with a cloud data center without sharing raw data. This fully utilizes global data to improve the model's generalization ability. Simultaneously, the continuous learning mechanism enables the model to continuously learn new fault modes and load change patterns, further enhancing its adaptability and robustness. A device fault diagnosis model constructed using deep learning algorithms effectively improves the ability to extract fault features, thereby increasing diagnostic accuracy. Furthermore, the automatic feature extraction capability of the deep learning model reduces manual intervention and improves diagnostic efficiency. An energy dispatching model constructed using swarm intelligence optimization algorithms dynamically adjusts the energy dispatching scheme based on real-time fault prediction results and monitoring data. This effectively handles multi-objective optimization problems, maximizing the use of renewable energy and reducing operating costs while ensuring the safe and stable operation of the system. Finally, a load forecasting model constructed using a fusion model deep learning algorithm effectively improves prediction accuracy and robustness by fusing multiple models, reducing the limitations of a single model.

[0015] Other beneficial effects of the present invention will be further explained in the specific embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart of the load prediction method for off-grid microgrid systems based on neural networks in this invention.

[0017] Figure 2 This is a structural block diagram of the off-grid microgrid system load prediction system based on neural networks in this invention. Detailed Implementation

[0018] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1: like Figure 1 As shown, this embodiment provides a load prediction method for off-grid microgrid systems based on neural networks, including the following steps: S1: Using neural network algorithms, construct equipment fault diagnosis models, energy dispatch models, and load forecasting models for off-grid microgrid systems, including the following steps: S1-1: Establish an edge-cloud collaborative architecture for an off-grid microgrid system; the edge-cloud collaborative architecture includes a cloud data center, regional base stations for several monitoring areas, and several edge computing gateways for each monitoring area. S1-2: Based on cloud data centers, use deep learning algorithms to build initial equipment fault diagnosis models and initial load prediction models, and use swarm intelligence optimization algorithms to build initial energy scheduling models. The device fault diagnosis model is constructed based on a Convolutional Neural Networks (CNN)-Long Short-Term Memory (LSTM) algorithm, and the device fault diagnosis model comprises a matrix feature extraction module and a device fault diagnosis module connected in sequence, wherein the matrix feature extraction module is constructed based on a CNN algorithm, and the device fault diagnosis module is constructed based on an LSTM algorithm. The matrix feature extraction module utilizes the powerful feature extraction capability of the CNN to perform feature extraction on the input matrix data. The CNN can automatically identify local features in the data through its unique convolution and pooling operations, and convert these features into higher-level abstract representations. This deep neural network structure can automatically extract features related to faults from the original data without the need for manual feature engineering, which not only improves the diagnostic accuracy of the model, but also greatly reduces the need for human intervention, making the model more intelligent and automated. The device fault diagnosis module utilizes the time series processing capability of the LSTM to further analyze the feature sequence output by the matrix feature extraction module. The LSTM can effectively capture long-term dependencies in time series data, thereby identifying the development trend and pattern of device faults. In an off-grid microgrid system, this helps to predict the future operating state of the device and timely detect potential fault risks. The load prediction model is constructed based on a Graph Convolutional Network (GCN)-Random Forest (RF)-Deep Belief Network (DBN)-SVM algorithm, and the load prediction model comprises a graph structure feature extraction module constructed based on a GCN algorithm, a key feature screening module constructed based on a RF algorithm, a load prediction module constructed based on a DBN algorithm, and an error compensation prediction module constructed based on a SVM algorithm. The graph structure feature extraction module, the key feature screening module, and the load prediction module are connected in sequence, and the error compensation prediction module is connected with the key feature screening module. The graph structure feature extraction module utilizes the GCN algorithm to extract the graph structure features of the input data. The GCN can effectively process graph structure data, capture relationship information between nodes (monitoring areas), and extract features related to load prediction. Through graph structure feature extraction, the model can better understand the mutual relationships between various monitoring areas in the off-grid microgrid system, thereby improving the accuracy of load prediction. The key feature screening module screens the extracted features using a random forest (RF) algorithm to identify the most critical features for load prediction. The RF algorithm builds multiple decision trees and performs ensemble learning to effectively evaluate the importance of features. Key feature screening reduces model complexity, improves computational efficiency, and retains the most important information for load prediction, thereby improving prediction performance. The load prediction module uses a deep belief network (DBN) algorithm for load prediction. DBN is a deep learning model that can learn complex nonlinear relationships from large amounts of data and be used to predict future load changes. Through deep learning-based load prediction, the model can accurately predict the future load demand of the off-grid microgrid system, providing strong support for energy scheduling. The error compensation prediction module uses a support vector machine (SVM) algorithm to compensate for prediction errors. SVM algorithm can effectively handle high-dimensional data and has strong generalization ability for predicting error values. The energy scheduling model is constructed based on the Fast Sparrow Search Algorithm (FSSA) algorithm, and the energy scheduling model includes the influence factor generation module, the fitness function update module, the initial solution generation module, the iteration update module, and the solution vector analysis module connected in sequence. S1-3: According to the continuous learning mechanism, set the initial device fault diagnosis model, the initial load prediction model, and the initial energy scheduling model experience replay pool and comprehensive loss function; S1-4: Based on the edge-cloud collaborative architecture, according to the federated learning mechanism, training optimization is performed to obtain the final device fault diagnosis model, the final load prediction model, and the final energy scheduling model, including the following steps: S1-4-1: Extract the model metadata of the initial device fault diagnosis model, the initial load prediction model, and the initial energy scheduling model, and send the model metadata to all regional base stations of the edge-cloud collaborative architecture; S1-4-2: Based on the regional base station, according to the model metadata, complete the reconstruction of the initial device fault diagnosis model, the initial load prediction model, and the initial energy scheduling model; S1-4-3: According to the collected historical regional monitoring data, perform optimization training to obtain the optimized device fault diagnosis model, the optimized load prediction model, the optimized energy scheduling model, and the historical experience; The historical system monitoring data includes a plurality of historical regional monitoring data, and the historical regional monitoring data includes a plurality of historical device monitoring data, each historical device monitoring data includes historical device dynamic monitoring data and historical device static rated data of the corresponding microgrid device. The historical device dynamic monitoring data includes historical meteorological condition monitoring data, historical user power consumption habit monitoring data, historical energy storage device state monitoring data, historical distributed power output power monitoring data, and historical device operation state monitoring data, and the historical device static rating data includes historical energy storage device rated energy storage parameters, historical distributed power rated output power, and historical load access rated parameters; S1-4-4: Extract the first model parameters of the optimized training device fault diagnosis model, the optimized training load prediction model, and the optimized training energy scheduling model, and upload the first model parameters and a plurality of historical experiences to the cloud data center; S1-4-5: Based on the cloud data center, integrate all the first model parameters uploaded by the regional base stations according to the alliance learning mechanism to obtain comprehensive model parameters, and store the plurality of historical experiences uploaded by all the regional base stations to the experience playback pool; S1-4-6: Adjust the initial device fault diagnosis model, the initial load prediction model, and the initial energy scheduling model according to the comprehensive model parameters to obtain the final device fault diagnosis model, the final load prediction model, and the final energy scheduling model; S1-4-7: Extract the second model parameters of the final device fault diagnosis model and the final energy scheduling model, and send the second model parameters to all the regional base stations to complete the synchronization of the final device fault diagnosis model and the final energy scheduling model; S2: Collect data and preprocess the real-time regional monitoring data of the off-grid micro-grid system in each monitoring area to obtain preprocessed real-time regional monitoring data, including the following steps: S2-1: Based on the edge computing gateway, collect heterogeneous real-time device monitoring data of different data sources in each monitoring area of the off-grid micro-grid system; S2-2: According to the preset data format, perform data mapping and conversion on a plurality of heterogeneous real-time device monitoring data to obtain a plurality of homogeneous real-time device monitoring data; S2-3: Perform data cleaning, Gaussian denoising, and normalization processing on a plurality of homogeneous real-time device monitoring data in sequence to obtain a plurality of standard real-time device monitoring data; S2-4: Perform data dimensionality reduction on a plurality of standard real-time device monitoring data to obtain a plurality of reduced real-time device monitoring data, and integrate to obtain preprocessed real-time regional monitoring data; The real-time system monitoring data includes a plurality of real-time regional monitoring data, and each real-time regional monitoring data includes a plurality of real-time device monitoring data, and each real-time device monitoring data includes real-time device dynamic monitoring data and real-time device static rating data of the corresponding micro-grid device; The real-time device dynamic monitoring data includes real-time meteorological condition monitoring data, real-time user electricity consumption habit monitoring data, real-time energy storage device state monitoring data, real-time distributed power output power monitoring data and real-time device operation state monitoring data, and the real-time device static rating data includes real-time energy storage device rated energy storage parameters, real-time distributed power rated output power and real-time load access rated parameters; S3: According to the pre-processed real-time regional monitoring data, using the device fault diagnosis model, device fault diagnosis is performed to obtain real-time device fault diagnosis results, including the following steps: S3-1: Based on the regional base station, the pre-processed real-time regional monitoring data including several reduced real-time device monitoring data is converted into a real-time regional monitoring data matrix; S3-2: Using the matrix feature extraction module of the device fault diagnosis model, the real-time matrix features of the real-time regional monitoring data matrix are extracted; S3-3: Using the device fault diagnosis module of the device fault diagnosis model, according to the real-time matrix features, device fault diagnosis is performed to obtain real-time device fault diagnosis results, and the first real-time experience segment generated by the device fault diagnosis model in the process of performing device fault diagnosis on the pre-processed real-time regional monitoring data is extracted; S3-4: If the real-time device fault diagnosis result is that there is a fault, the energy scheduling step is entered, otherwise all pre-processed real-time regional monitoring data is integrated to obtain real-time system monitoring data, and the load prediction step is entered; Specifically, the pre-processed real-time regional monitoring data of all monitoring regions is integrated, and according to the power connection cable lines between each monitoring region and other monitoring regions in the off-grid micro-grid system, the graph structure of the pre-processed real-time regional monitoring data is converted to obtain the graph structure of the real-time system monitoring data; S4: If the real-time device fault diagnosis result is that there is a fault, the energy scheduling step is entered, otherwise all pre-processed real-time regional monitoring data is integrated to obtain real-time system monitoring data, and the load prediction step is entered; S5: According to the real-time device fault diagnosis result and the pre-processed real-time regional monitoring data, using the energy scheduling model, energy scheduling is performed, the obtained real-time energy scheduling scheme is executed, and the data collection step is returned, including the following steps: S5-1: The regional base station uses the influence factor generation module of the energy scheduling model to analyze the real-time device fault diagnosis result to obtain real-time influence factors, and sets real-time optimization targets according to the real-time influence factors; Taking the real-time device fault diagnosis result that there is an energy storage device fault as an example, the real-time influence factors are the energy storage device fault influence range and the energy storage device maintenance cost, and the real-time optimization target is to minimize the energy storage device fault influence range and the energy storage device maintenance cost; S5-2: Using the fitness function update module of the energy dispatch model, the preset fitness function of the energy dispatch model is updated according to the real-time optimization objective to obtain the real-time fitness function; The formula for the real-time fitness function is:

[0020] In the formula, For FSSA individuals The real-time fitness value; The scope of impact of an energy storage device failure; For the maintenance cost of energy storage devices; For FSSA individual variables; c For FSSA individual indicators; These are the first and second weight values; S5-3: Using the initial solution generation module of the energy dispatch model, based on the real-time equipment fault diagnosis results and preprocessed real-time regional monitoring data, set the real-time energy dispatch scheme format and real-time constraints; S5-4: Using the initial solution generation module of the energy dispatch model, an initial solution is generated according to the real-time energy dispatch scheme format and real-time constraints, resulting in several initial solutions; the initial solutions correspond to the initial real-time energy dispatch scheme. The formula is:

[0021] In the formula, The initial FSSA individuals for the Circle chaotic mapping; The initial FSSA individuals are randomly generated; c For FSSA individual indicators; S5-5: Using the iterative update module of the energy dispatch model, based on the real-time fitness function and real-time constraints, iteratively updates several initial solutions and outputs the optimal solution with the best real-time fitness value, including the following steps: S5-5-1: Use the real-time fitness function to obtain the fitness value of each initial FSSA individual (initial solution) in the initial FSSA population; S5-5-2: Based on the fitness values ​​of the initial FSSA individuals, sort the initial FSSA individuals to obtain the initial discoverer, the initial joiner, and the initial predator. S5-5-3: Update the initial FSSA population to obtain an updated FSSA population; the updated FSSA population includes updated discoverers, updated joiners, and updated predators. The update formula for the discoverer is:

[0022] where, are the first t +1, t th discoverer FSSA individuals in the c th iteration; is the maximum iteration number; is a random number between 0 and 1; is a normally distributed random number; is a matrix with all elements being 1; is an alert value; is a security threshold; is a matrix parameter; D The update formula of the joiner is:

[0023] where, are the first t +1, t th joiner FSSA individuals in the c th iteration; is the best position occupied by the exposed identity; is the current worst position; is a random number between 0 and 1; is a matrix with all elements being 1 or -1; is a sparrow indicator; c is the total number of FSSA individuals; h The update formula of the predator is:

[0024] where, are the first t +1, t th predator FSSA individuals in the c th iteration; is a step control parameter, and , is a convergence factor, is a positive step control real number not equal to 0; is the current best position; are the current, best, and worst fitness of the FSSA individuals, respectively; is a minimum constant to prevent the denominator from being 0;

[0025] where, is a convergence factor; tanh(.) is the hyperbolic tangent function; a ​​max , a min are respectively the maximum and minimum values of the convergence factor; λ is a decreasing rate parameter, is a decreasing period parameter, λ = -2 π , π ; S5-5-4: using a dynamic reverse learning algorithm, performing dynamic reverse learning on the updated FSSA population to generate a dynamic reverse FSSA population; The formula is:

[0026] In the formula, is a dynamic reverse FSSA individual; γ is a decreasing inertia coefficient; is an upper limit of the search space; is a lower limit of the search space; is an updated FSSA individual; S5-5-5: according to the fitness function, calculating the fitness values of all FSSA individuals in the updated FSSA population and the dynamic reverse FSSA population, and taking the FSSA individual with the minimum fitness value as the optimal individual; S5-5-6: if the number of optimization iterations reaches the maximum number of iterations or the fitness value of the optimal individual meets the requirements, output the optimal solution; S5-6: using the solution vector analysis module of the energy scheduling model, performing solution vector analysis on the individual vector of the optimal solution to obtain the optimal real-time energy scheduling scheme; S5-7: extracting the second real-time experience segment generated by the energy scheduling model in the process of scheduling energy after real-time device fault diagnosis results and preprocessed real-time regional monitoring data, and uploading it to the cloud data center; S5-7: publishing the optimal real-time energy scheduling scheme to the off-grid micro-grid system, executing the optimal real-time energy scheduling scheme, and returning to the data collection step; S6: using the load prediction model, performing load prediction according to the real-time system monitoring data of the off-grid micro-grid system to obtain real-time load prediction results, including the following steps: S6-1: using the graph structure feature extraction module of the load prediction model, extracting real-time graph structure features of the real-time system monitoring data; S6-2: using the key feature screening module of the load prediction model, extracting several real-time key features of the real-time graph structure features; ​S6-3: The load prediction module using the load prediction model performs load prediction diagnosis based on several real-time key features to obtain the real-time load prediction probability distribution. S6-4: The error compensation prediction module using the load prediction model performs error compensation prediction based on several real-time key features to obtain real-time error compensation values. S6-5: Add the real-time load prediction probability distribution and the corresponding real-time error compensation value to obtain the real-time load prediction result of the off-grid microgrid system; S6-6: Perform updates to obtain updated equipment fault diagnosis models, updated energy dispatch models, and updated load prediction models, including the following steps: S6-6-1: Extract the third real-time experience segment generated by the load prediction model during the load prediction process of real-time system monitoring data, and combine it with the first real-time experience segment and / or the second real-time experience segment uploaded by base stations in several regions to obtain several real-time experiences. S6-6-2: Randomly extract several historical experiences from the experience replay pool and mix them with several real-time experiences to obtain several mixed experiences; S6-6-3: Based on several hybrid experiences, the equipment fault diagnosis model, energy dispatch model and load prediction model are continuously trained, and the real-time loss value generated by each continuous training is extracted using a comprehensive loss function. S6-6-4: If the real-time loss value is less than the loss value threshold or the number of continuous training sessions reaches the number threshold, then the updated equipment fault diagnosis model, the updated energy scheduling model, and the updated load prediction model are obtained.

[0027] Example 2: like Figure 2 As shown, this embodiment provides a load prediction system for off-grid microgrid systems based on neural networks, used to implement a load prediction method for off-grid microgrid systems. The system includes a cloud data center, several regional base stations, several edge computing gateways, several monitoring data acquisition devices, and several microgrid devices. The cloud data center is communicatively connected to the regional base stations of several monitoring areas. Each regional base station is communicatively connected to several edge computing gateways of the corresponding monitoring area. Each edge computing gateway is communicatively connected to several monitoring data acquisition devices. Each set of several monitoring data acquisition devices is communicatively connected to the corresponding microgrid devices.

[0028] The cloud data center is used to construct equipment fault diagnosis models, energy dispatch models, and load forecasting models for off-grid microgrid systems using neural network algorithms; based on real-time system monitoring data of the off-grid microgrid system, the load forecasting model is used to perform load forecasting and obtain real-time load forecasting results. The regional base station is used for carrying out equipment fault diagnosis by using the equipment fault diagnosis model according to the preprocessed real-time regional monitoring data, obtaining real-time equipment fault diagnosis results, integrating all the preprocessed real-time regional monitoring data, obtaining real-time system monitoring data, using the energy scheduling model according to the real-time equipment fault diagnosis results and the preprocessed real-time regional monitoring data, carrying out energy scheduling, executing the obtained real-time energy scheduling scheme, and returning to the data collection step; The edge computing gateway is used for collecting data, and pre-processing real-time equipment monitoring data of micro-grid equipment in each monitoring area of the off-grid micro-grid system to obtain pre-processed real-time regional monitoring data. The monitoring data collection device is used for collecting real-time equipment monitoring data of corresponding micro-grid equipment and sending the real-time equipment monitoring data to the edge computing gateway of the corresponding monitoring area.

[0029] The off-grid micro-grid system load prediction method and system based on a neural network provided by the application adopt a coalition learning mechanism, each regional base station can perform collaborative training by exchanging model parameters with a cloud data center without sharing original data, fully utilize global data to improve model generalization ability, and a continuous learning mechanism enables the model to continuously learn new fault modes and load change rules, further improving the adaptability and robustness of the model; the equipment fault diagnosis model constructed by using a deep learning algorithm effectively improves the extraction ability of fault features, thereby improving the diagnosis accuracy, and the automatic feature extraction ability of the deep learning model reduces manual intervention and improves the diagnosis efficiency; the energy scheduling model constructed by using a swarm intelligence optimization algorithm dynamically adjusts the energy scheduling scheme according to real-time fault prediction results and monitoring data; can effectively handle multi-objective optimization problems, maximize the use of renewable energy under the premise of meeting the safe and stable operation of the system, and reduce operation costs; the load prediction model constructed by using a fusion model deep learning algorithm can effectively improve the prediction accuracy and robustness, and reduce the limitations of a single model.

[0030] The application is not limited to the above-mentioned optional embodiments, and anyone can derive other various forms of products under the inspiration of the application. The above-mentioned specific embodiments should not be understood as limiting the protection scope of the application, and the protection scope of the application should be defined by the claims, and the specification can be used to explain the claims.

Claims

1. A neural network-based load forecasting method for an off-grid microgrid system, characterized in that: The method comprises the following steps: Using a neural network algorithm, a device fault diagnosis model, an energy scheduling model, and a load prediction model of an off-grid micro-grid system are constructed. Data collection is performed, and real-time regional monitoring data of the off-grid micro-grid system in each monitoring area is preprocessed to obtain preprocessed real-time regional monitoring data. According to the preprocessed real-time regional monitoring data, a device fault diagnosis model is used to perform device fault diagnosis to obtain real-time device fault diagnosis results. If the real-time device fault diagnosis result is a fault, the energy scheduling step is entered, otherwise all preprocessed real-time regional monitoring data is integrated to obtain real-time system monitoring data, and the load prediction step is entered. According to the real-time device fault diagnosis result and the preprocessed real-time regional monitoring data, an energy scheduling model is used to perform energy scheduling, and the obtained real-time energy scheduling scheme is executed, and the data collection step is returned. According to the real-time system monitoring data of the off-grid micro-grid system, a load prediction model is used to perform load prediction to obtain real-time load prediction results. 2.The neural network-based load forecasting method for off-grid microgrid system according to claim 1, wherein: Using a neural network algorithm, a device fault diagnosis model, an energy scheduling model, and a load prediction model of an off-grid micro-grid system are constructed, comprising the following steps: A cloud-edge collaborative architecture of the off-grid micro-grid system is built; the cloud-edge collaborative architecture comprises a cloud data center, regional base stations of a plurality of monitoring areas, and a plurality of edge computing gateways of each monitoring area; Based on the cloud data center, an initial device fault diagnosis model and an initial load prediction model are constructed using a deep learning algorithm, and an initial energy scheduling model is constructed using a swarm intelligence optimization algorithm; According to a continuous learning mechanism, an experience replay pool and a comprehensive loss function of the initial device fault diagnosis model, the initial load prediction model, and the initial energy scheduling model are set; Based on the cloud-edge collaborative architecture, the final device fault diagnosis model, the final load prediction model, and the final energy scheduling model are obtained through training and optimization according to a federated learning mechanism.

3. The neural network-based load forecasting method for off-grid microgrid systems according to claim 2, wherein: The device fault diagnosis model is constructed based on a CNN-LSTM algorithm.

4. The neural network-based load forecasting method for off-grid microgrid systems according to claim 3, wherein: The load prediction model is constructed based on a GCN-RF-DBN-SVM algorithm.

5. The neural network-based load forecasting method for off-grid microgrid systems according to claim 4, wherein: The energy scheduling model is constructed based on a FSSA algorithm.

6. The neural network-based load forecasting method for off-grid microgrid systems according to claim 5, wherein: Data collection is performed, and real-time regional monitoring data of the off-grid micro-grid system in each monitoring area is preprocessed to obtain preprocessed real-time regional monitoring data, comprising the following steps: Heterogeneous real-time device monitoring data of different data sources in each monitoring area of the off-grid micro-grid system is collected; According to a preset data format, a plurality of heterogeneous real-time device monitoring data is mapped and converted to obtain a plurality of homogeneous real-time device monitoring data; A plurality of homogeneous real-time device monitoring data is sequentially subjected to data cleaning, Gaussian denoising, and normalization processing to obtain a plurality of standard real-time device monitoring data; A plurality of standard real-time device monitoring data is subjected to data dimension reduction to obtain a plurality of reduced real-time device monitoring data, and the preprocessed real-time regional monitoring data is obtained by integration.

7. The neural network-based load forecasting method for off-grid microgrid systems according to claim 6, wherein: According to the pre-processed real-time regional monitoring data, using the equipment fault diagnosis model, equipment fault diagnosis is carried out, and real-time equipment fault diagnosis results are obtained, including the following steps: The pre-processed real-time regional monitoring data including a plurality of reduced real-time equipment monitoring data is converted into a real-time regional monitoring data matrix; Using the equipment fault diagnosis model, the real-time matrix features of the real-time regional monitoring data matrix are extracted; According to the real-time matrix features, equipment fault diagnosis is carried out, and real-time equipment fault diagnosis results are obtained; If the real-time equipment fault diagnosis result is a fault, the energy scheduling step is entered, otherwise all pre-processed real-time regional monitoring data is integrated to obtain real-time system monitoring data, and the load prediction step is entered.

8. The neural network-based load forecasting method for off-grid microgrid systems according to claim 7, wherein: According to the real-time equipment fault diagnosis result and the pre-processed real-time regional monitoring data, using the energy scheduling model, energy scheduling is carried out, the obtained real-time energy scheduling scheme is executed, and the data collection step is returned, including the following steps: Using the energy scheduling model, the real-time influence factor is obtained by analyzing the real-time equipment fault diagnosis result, and the real-time optimization target is set according to the real-time influence factor; According to the real-time optimization target, the pre-set fitness function of the energy scheduling model is updated to obtain a real-time fitness function; According to the real-time equipment fault diagnosis result and the pre-processed real-time regional monitoring data, the real-time energy scheduling scheme format and the real-time constraint condition are set; According to the real-time energy scheduling scheme format and the real-time constraint condition, initial solutions are generated to obtain a plurality of initial solutions; The initial solution corresponds to an initial real-time energy scheduling scheme; According to the real-time fitness function and the real-time constraint condition, the plurality of initial solutions are iteratively updated, and the optimal solution with the optimal real-time fitness value is output; The individual vector of the optimal solution is analyzed by vector analysis to obtain the optimal real-time energy scheduling scheme; The optimal real-time energy scheduling scheme is published to the off-grid micro-grid system, the optimal real-time energy scheduling scheme is executed, and the data collection step is returned.

9. The neural network-based load forecasting method for off-grid microgrid systems according to claim 8, wherein: According to the real-time system monitoring data of the off-grid micro-grid system, using the load prediction model, load prediction is carried out, and real-time load prediction results are obtained, including the following steps: Integrate all pre-processed real-time regional monitoring data of the monitoring area to obtain real-time system monitoring data of the graph structure; Using the load prediction model, the real-time graph structure features of the real-time system monitoring data are extracted; A plurality of real-time key features of the real-time graph structure features are extracted; According to the plurality of real-time key features, load prediction diagnosis is carried out to obtain a real-time load prediction probability distribution; According to the plurality of real-time key features, error compensation prediction is carried out to obtain a real-time error compensation value; The real-time load prediction probability distribution and the corresponding real-time error compensation value are added to obtain the real-time load prediction result of the off-grid micro-grid system; Update to obtain an updated equipment fault diagnosis model, an updated energy scheduling model and an updated load prediction model.

10. A neural network based off-grid microgrid system load forecasting system for implementing the off-grid microgrid system load forecasting method according to any one of claims 1-9, characterized in that: The system comprises a cloud data center, a plurality of regional base stations, a plurality of edge computing gateways, a plurality of monitoring data collection devices and a plurality of micro-grid devices, the cloud data center is in communication connection with the regional base stations of a plurality of monitoring areas respectively, each regional base station is in communication connection with the edge computing gateways of the corresponding monitoring area respectively, each edge computing gateway is in communication connection with the monitoring data collection devices, and each monitoring data collection device is in communication connection with the corresponding micro-grid device.

Images