Multi-source power distribution network energy-saving optimization scheduling method, system, equipment and storage medium based on big data processing

Abstract

The invention discloses a multi-source power distribution network energy-saving optimization scheduling method, system, equipment and storage medium based on big data processing, and relates to the technical field of power systems and big data processing. After power distribution network multi-source data is collected, through preprocessing and feature extraction, loads are predicted and classified through a long and short-term memory network; the load dynamic optimization module combines cost, network loss and photovoltaic consumption, generates a scheduling instruction through a mixed integer programming model, compares power loss before and after scheduling after CPLEX solution, and generates a difference thermodynamic diagram. According to the scheme, accurate load prediction and dynamic optimization scheduling of the multi-source power distribution network are realized through big data processing, the cost and energy efficiency are effectively improved, the consumption capability of photovoltaic energy is enhanced, the energy-saving effect is visually displayed through the power loss difference thermodynamic diagram before and after scheduling, and powerful support is provided for intelligent management of the power distribution network.

Description

Technical Field

[0001] The present invention relates to the field of power systems and big data processing technology, and specifically to a multi-source distribution network energy-saving optimization scheduling method, system, equipment, and storage medium based on big data processing. Background Art

[0002] A multi-source distribution network is a distribution system that can receive electricity from multiple power sources. These power sources may include traditional power plants, transmission networks, and emerging distributed energy resources such as solar and wind power generation facilities. The system uses advanced intelligent scheduling and control technologies to achieve optimal combination and complementarity between different power sources, thereby improving the reliability and flexibility of power supply. The advantage of a multi-source distribution network is that when a single power source fails, other power sources can quickly supplement power supply, effectively reducing the risk of power outages. At the same time, the system can also reduce energy loss and improve energy utilization efficiency by optimizing power source combination and scheduling. With the rapid development of renewable energy and distributed energy resources, and the continuous advancement of smart grid technology, multi-source distribution networks are gradually becoming an important trend in the development of power systems. They can better adapt to the needs of diversified energy structures and meet users' requirements for power quality, power supply reliability, and environmental protection.

[0003] In order to solve the problem of multi-objective coordinated optimization of cost, energy efficiency and photovoltaic absorption in multi-source distribution networks, existing technologies mainly rely on traditional optimization algorithms and models for prediction and scheduling. However, this method is powerless when processing large-scale, high-dimensional data, and often results in low computational efficiency and inaccurate optimization results. Especially in the face of real-time changes in photovoltaic power generation and complex grid load demands, traditional methods find it difficult to respond quickly and make optimal scheduling decisions, which in turn leads to increased grid operating costs, reduced energy efficiency and waste of photovoltaic energy. In order to overcome these challenges, a multi-source distribution network energy-saving optimization scheduling method, system, equipment and storage medium based on big data processing are proposed. Summary of the Invention

[0004] The object of the present invention is to provide a multi-source distribution network energy-saving optimization scheduling method, system, equipment, and storage medium based on big data processing, through multi-source data collection, preprocessing, feature extraction, load forecasting classification, load dynamic optimization and energy efficiency scheduling analysis, wherein smart meters are used to collect active power and photovoltaic inverter monitoring devices are used to collect photovoltaic power generation power. After operations such as outlier processing and data normalization, a long-short-term memory network model is constructed to predict load, and scheduling is optimized based on a particle swarm algorithm and a mixed integer programming model. The CPLEX solver is called to generate scheduling instructions and draw a difference heat map to achieve energy-saving optimization scheduling and continuous and efficient operation of the distribution network, so as to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: A multi-source distribution network energy-saving optimization scheduling method based on big data processing, characterized by comprising the following steps: S1: Multi-source data collection of distribution network, including active power collection of distribution network and photovoltaic power generation power collection; S2: Distribution network data preprocessing, including outlier detection and removal, data normalization, and noise filtering; S3: Data feature extraction, including active power fluctuation feature extraction, photovoltaic power generation power fluctuation pattern recognition and correlation analysis; S4: Load forecast classification, by building a load demand forecast model and classifying the load state through the load demand forecast model prediction; S5: Dynamic load optimization, through state-weight mapping, adjusting particle swarm inertia weight, and building a photovoltaic absorption constraint model; S6: Energy efficiency scheduling analysis, generating scheduling instructions and generating difference heat maps.

[0006] Step S1: collecting multi-source data of the distribution network, including collecting active power of the distribution network and collecting photovoltaic power generation power; The distribution network active power collection is connected to the three-phase power line through the distribution network node, and connected to the regional data concentrator through the power line carrier. The smart meter is configured with an independent device code, mapped to the multi-source distribution network energy-saving optimization scheduling system database according to the node topology relationship, and the line current is sensed in real time by the built-in current transformer of the smart meter to calculate the active power. The photovoltaic power generation power collection is deployed on the DC side and AC side of the grid-connected inverter of the photovoltaic power station through a photovoltaic inverter monitoring device. The DC side uses a Hall sensor to measure the output current of the photovoltaic string, and the maximum power point is calculated in combination with the open-circuit voltage of the photovoltaic string to output the photovoltaic power generation power.

[0007] Step S2: pre-processing the distribution network data, specifically including the following steps: S21: Outlier detection and elimination: Receive data streams from smart meters and photovoltaic inverters, set the sliding window width, calculate the mean and standard deviation of active power and photovoltaic power generation within the window point by point, and determine whether the current data point deviates from the mean by more than 3 times the standard deviation. If so, it is marked as an outlier and eliminated from the data stream; if not, it is marked as a normal value. In view of the fluctuation characteristics of photovoltaic power generation, an adaptive threshold is introduced during anomaly detection. Specifically, a sliding window is set, the width of which is determined according to the time resolution and fluctuation characteristics of the data. The sliding window is used to calculate the mean and standard deviation of active power and photovoltaic power generation within the window point by point. Within the sliding window, the standard deviation of the active power and photovoltaic power generation data is calculated respectively. Then, the standard deviation within the sliding window is multiplied by a dynamic coefficient to obtain the adaptive threshold. S22: Data normalization: Calculate the minimum and maximum values ​​of active power and photovoltaic power generation within a preset time period. Use a linear transformation formula to map the original data to the [0, 1] interval. Update the minimum and maximum values ​​every hour. S23: Noise filtering: For power data with obvious high-frequency noise interference, a sliding average algorithm with a window width of 5 seconds is used to calculate the mean of the data in the window as the output, and the maximum change slope of the data in the filtering window is limited.

[0008] Step S3 data feature extraction specifically includes the following steps: S31: Extract active power fluctuation characteristics, including power fluctuation amplitude and load peak-valley fluctuation range; using a fixed time period as a statistical window, take the difference between the maximum and minimum values ​​of the normalized active power data within the window, divide it by the time period length, and obtain the power fluctuation amplitude per unit time, which is used to judge the stability of the power grid; within a preset period, calculate the difference between the global maximum and minimum values ​​of the normalized active power data as the peak-valley difference characteristic value to quantify the peak-valley fluctuation range of the load and identify the load characteristics of the power grid; S32: Photovoltaic power fluctuation pattern identification: Through power spectrum density analysis, the main fluctuation frequency components in photovoltaic power data are extracted, their energy proportions are calculated, and the periodic fluctuation patterns of output are identified based on their energy proportions. This is used to build a photovoltaic consumption constraint model and optimize the scheduling strategy. S33: Correlation analysis, using sliding windows as units, calculate the covariance of photovoltaic power generation and active power, and divide them by the product of their standard deviations to obtain the photovoltaic-load Pearson correlation coefficient; if the correlation coefficient is positive, it means that the photovoltaic output and load demand change synchronously; if it is negative, it means inverse fluctuation. According to the topological relationship, calculate the correlation coefficient between the active power of each node. Specifically, calculate the Pearson correlation coefficient between the active power of each node, process the active power data of each node through a sliding window method, calculate the covariance between each pair of nodes, and divide them by the product of their standard deviations to obtain the correlation coefficient, and generate an adjacency matrix through the correlation coefficient. Use the generated adjacency matrix to analyze the correlation distribution and connection pattern between the nodes. Through the matrix analysis method, identify the load propagation path in the distribution network, that is, determine the transfer of load from one node to other nodes, so as to identify the load propagation path and regional coupling characteristics, predict the load change trend and distribution to optimize the scheduling of distributed energy output and load demand.

[0009] Step S4 load forecast classification specifically includes the following steps: S41: Construct a load demand forecasting model, organize the pre-processed normalized active power data into sliding window data of historical periods according to time series, and form multi-dimensional time series input vector samples at fixed intervals. The samples contain active power values, node codes, and data quality flags with continuous time stamps, and are aligned into a space-time matrix according to the distribution network topology. A long-short-term memory network structure is constructed, consisting of an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is consistent with the historical time step, and it receives the active power sequence. The hidden layer selectively memorizes historical information and updates the cell state through a triple gating mechanism consisting of a forget gate, an input gate, and an output gate. The output layer is mapped to the load demand sequence for the future period, and the load demand sequence dimension matches the prediction duration. S42: Load demand forecasting and load state classification. The latest collected active power data is normalized and input into the trained load demand forecasting model. The historical information is filtered through the forgetting gate, the memory unit is updated by the input gate, and the hidden state is generated by the output gate to output the load demand sequence forecast result. The maximum value is extracted from the load demand forecast result as the predicted load peak, and the predicted load peak is divided by the rated total capacity of the distribution network and multiplied by 100% to obtain the predicted load rate in percentage form. The predicted load rate range is determined and the corresponding load state label is generated, including high load state, normal load state and low load state, and the load state label is bound to the load rate value. If the predicted load rate is ≥85%, it is a high load state. If it is 60%≤predicted load rate<85%, it is a normal load state. If the predicted load rate is <60%, it is a low load state. A recursive prediction strategy is used to balance the long-term prediction accuracy and error accumulation problems.

[0010] Step S5 load dynamic optimization classifies the load status in the previous step, generates high load, normal load and low load state label inputs, and drives the load dynamic optimization module. Specifically, it includes the following steps: S51: State-weight mapping: Construct a state-weight mapping table. If the system is in a high-load state, the objective function focuses on reducing network loss; if the system is in a normal-load state, the objective function focuses on balancing optimization cost and network loss; if the system is in a low-load state, the objective function focuses on cost. Receive the load status tag, search the status-weight mapping table in real time to obtain the preset weight value, and trigger a smooth weight transition if the status tags are the same for two consecutive cycles; S52: Adjust the particle swarm inertia weight, define the economic cost, network loss cost, and cost-network loss objective functions, and dynamically adjust the particle swarm inertia weight w according to the load state label; if it is in a high load state, accelerate the convergence speed to prioritize stability; if it is in a normal load state, set the balance between global search and local convergence capabilities; if it is in a low load state, set the enhanced global search capability to explore cost potential; The ratio of cost weight to network loss weight is used as the boundary condition for particle position update to limit the search space; S53: Construct a photovoltaic absorption constraint model based on a mixed integer programming model architecture. Define photovoltaic output, energy storage charging and discharging power, and main grid purchase power as continuous variables, and photovoltaic unit switching status and transformer tap position as integer variables. Convert the weight-driven load state-related photovoltaic absorption constraints into linear constraints. If the system is in a high-load state, the photovoltaic power generation power is limited to no more than 0.9 times the rated photovoltaic power generation power, and the energy storage is forced to discharge to make up for the shortfall; if the system is in a normal-load state, the photovoltaic output benchmark value is limited to 0.8 to 1.1 times the current load, and the energy storage charge and discharge power is limited to 50% of the rated value; if the system is in a low-load state, the photovoltaic power generation power is limited to no less than 0.95 times the load power, and energy storage is allowed to reversely supply power to the main grid.

[0011] Step S6: Energy efficiency scheduling analysis, specifically including the following steps: S61: Generates dispatch instructions, receives a PV consumption constraint model containing a load-state-driven objective function, continuous variables, and integer variables, converts the objective function and constraints into a mathematical format recognizable by the solver, calls the CPLEX solver, uses a branch-and-bound algorithm to traverse the feasible solution space, prioritizes integer variable combinations, calculates node currents and network losses, evaluates the objective function value, and selects a Pareto optimal solution set that satisfies the weight constraints. After the solution is completed, the specific values ​​of the continuous variables are extracted from the results returned by the solver. Based on the analyzed continuous variable values, specific dispatch instructions are generated. The dispatch instructions include: generating photovoltaic inverter power setting instructions, energy storage charging and discharging power instructions, main network power purchase plan and transformer gear adjustment instructions; each type of generated instruction is encapsulated into a standardized control frame according to node code, timestamp and execution priority, and the instructions are broadcast to photovoltaic power stations, energy storage controllers and load management terminals in a 1-second cycle. The deviation rate between the instructions and real-time data is checked. If the deviation exceeds 5%, the model is re-solved, and the dispatch strategy is synchronized with the energy-saving optimization dispatch system status of the multi-source distribution network; S62: Generate a difference heat map. Before the dispatch instruction is issued, collect the current and line resistance of each node in real time to calculate the node power loss. After the dispatch instruction is executed, calculate the node power loss and loss difference based on the optimized current value to generate a loss difference heat map. Use red areas to mark high-loss nodes with positive loss differences, and the color scale depth is proportional to the loss increment; use green areas to mark valid nodes with negative loss differences, and the color scale depth is proportional to the loss reduction; use gray areas to mark nodes with no obvious change in loss difference. Click a node in the loss difference heat map to display the scheduling instruction details of the node. It supports replaying the heat map changes before and after scheduling along the timeline to analyze the spatiotemporal effects of the strategy.

[0012] A multi-source distribution network energy-saving optimization and dispatching system based on big data processing, characterized by comprising a distribution network multi-source data acquisition module, a distribution network data preprocessing module, a load forecasting and classification module, a load dynamic optimization module and an energy efficiency dispatching and analysis module; The distribution network multi-source data acquisition module collects the active power data of each node in the distribution network and the photovoltaic power generation data of the photovoltaic power station; The load forecasting classification module builds a load demand forecasting model based on active power data, thereby predicting the load demand of the distribution network and classifying the load status by label; The load dynamic optimization module uses the adaptive weight coefficient mechanism in the improved particle swarm algorithm to optimize the load state analysis results after label classification, maps the load state classification results to cost weights and network loss weights, and constructs a photovoltaic absorption constraint model; The energy efficiency scheduling analysis module calls the CPLEX solver to solve the photovoltaic absorption constraint model, generates scheduling instructions, and generates a difference heat map based on the power loss of the distribution network before and after the scheduling instructions are taken effect.

[0013] A terminal device, characterized in that it includes a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the method according to any one of claims 1 to 7 is implemented.

[0014] A computer-readable storage medium, characterized in that a computer program is stored in the storage medium, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

[0015] Compared with the prior art, the present invention has the following beneficial effects: Improved data utilization and scheduling accuracy: Existing technologies rely on a single data source, limiting scheduling accuracy. This solution uses a multi-source data acquisition module to comprehensively collect active power data from each node in the distribution network and generated power data from photovoltaic power plants. After preprocessing and feature extraction, it greatly improves data integrity and reliability, providing a solid foundation for precise scheduling. For example, when forecasting load demand, processing multi-source data based on long-short-term memory networks can more accurately grasp load change trends. Compared with traditional methods, prediction errors are significantly reduced, making scheduling decisions more in line with actual needs and reducing imbalances in power supply and demand.

[0016] Optimizing Energy Savings: Traditional scheduling methods struggle to globally optimize energy-saving scheduling solutions. This invention leverages an improved particle swarm optimization algorithm and a mixed integer programming model, combining cost and network loss weighting optimization with photovoltaic consumption constraints to achieve multi-objective collaborative optimization. Under high load conditions, network loss reduction is prioritized; under normal load conditions, costs and network losses are balanced; under low load conditions, costs are prioritized while promoting renewable energy consumption. This significantly improves energy efficiency and promotes the green and sustainable development of the distribution network.

[0017] Enhanced real-time performance and responsiveness: Traditional dispatch systems have limited computing power, making them unable to process large amounts of data in real time and resulting in delayed decision-making. This system leverages big data preprocessing and feature extraction technologies to accelerate data analysis. Furthermore, it leverages the efficient CPLEX solver and optimization algorithms to rapidly calculate and generate dispatch instructions. This ensures that dispatch strategies are synchronized with the real-time status of the power grid, effectively improving the distribution network's ability to respond to emergencies and ensuring power supply stability and reliability.

[0018] Improved visualization and decision support: Existing technologies lack a means to visually demonstrate dispatch effects, hindering decision-making. This system generates differential heat maps to visually display the changes in power loss at each node before and after dispatch. Operations and maintenance personnel can clearly view the distribution of high-loss and effective nodes, click on a node to obtain detailed dispatch instructions, and replay the heat map changes over time, enabling in-depth analysis of the spatiotemporal effects of strategies. This provides strong support for optimizing dispatch strategies and enhancing intelligent management of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a diagram of a multi-source distribution network energy-saving optimization scheduling system based on big data processing according to the present invention.

[0020] Figure 2 This is a flow chart of a multi-source distribution network energy-saving optimization scheduling method based on big data processing according to the present invention. Specific implementation methods The technical solutions in the embodiments of the present invention will be fully described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0021] like Figure 1-2 As shown, a multi-source distribution network energy-saving optimization scheduling method based on big data processing is characterized by comprising the following steps: S1: Multi-source data collection of distribution network, including active power collection of distribution network and photovoltaic power generation power collection; S2: Distribution network data preprocessing, including outlier detection and removal, data normalization, and noise filtering; S3: Data feature extraction, including active power fluctuation feature extraction, photovoltaic power generation power fluctuation pattern recognition and correlation analysis; S4: Load forecast classification, by building a load demand forecast model and classifying the load state through the load demand forecast model prediction; S5: Dynamic load optimization, through state-weight mapping, adjusting particle swarm inertia weight, and building a photovoltaic absorption constraint model; S6: Energy efficiency scheduling analysis, generating scheduling instructions and generating difference heat maps.

[0022] Step S1: collecting multi-source data of the distribution network, including collecting active power of the distribution network and collecting photovoltaic power generation power; The distribution network active power collection is connected to the three-phase power line through the distribution network node, and connected to the regional data concentrator through the power line carrier. The smart meter is configured with an independent device code, mapped to the multi-source distribution network energy-saving optimization scheduling system database according to the node topology relationship, and the line current is sensed in real time by the built-in current transformer of the smart meter to calculate the active power. The photovoltaic power generation power collection is deployed on the DC side and AC side of the grid-connected inverter of the photovoltaic power station through a photovoltaic inverter monitoring device. The DC side uses a Hall sensor to measure the output current of the photovoltaic string, and the maximum power point is calculated in combination with the open-circuit voltage of the photovoltaic string to output the photovoltaic power generation power.

[0023] Step S2: pre-processing the distribution network data, specifically including the following steps: S21: Outlier detection and elimination: Receive data streams from smart meters and photovoltaic inverters, set the sliding window width, calculate the mean and standard deviation of active power and photovoltaic power generation within the window point by point, and determine whether the current data point deviates from the mean by more than 3 times the standard deviation. If so, it is marked as an outlier and eliminated from the data stream; if not, it is marked as a normal value. In view of the fluctuation characteristics of photovoltaic power generation, an adaptive threshold is introduced during anomaly detection. Specifically, a sliding window is set, the width of which is determined according to the time resolution and fluctuation characteristics of the data. The sliding window is used to calculate the mean and standard deviation of active power and photovoltaic power generation within the window point by point. Within the sliding window, the standard deviation of the active power and photovoltaic power generation data is calculated respectively. Then, the standard deviation within the sliding window is multiplied by a dynamic coefficient to obtain the adaptive threshold. S22: Data normalization: Calculate the minimum and maximum values ​​of active power and photovoltaic power generation within a preset time period. Use a linear transformation formula to map the original data to the [0, 1] interval. Update the minimum and maximum values ​​every hour. S23: Noise filtering: For power data with obvious high-frequency noise interference, a sliding average algorithm with a window width of 5 seconds is used to calculate the mean of the data in the window as the output, and the maximum change slope of the data in the filtering window is limited.

[0024] Step S3 data feature extraction specifically includes the following steps: S31: Extract active power fluctuation characteristics, including power fluctuation amplitude and load peak-valley fluctuation range; using a fixed time period as a statistical window, take the difference between the maximum and minimum values ​​of the normalized active power data within the window, divide it by the time period length, and obtain the power fluctuation amplitude per unit time, which is used to judge the stability of the power grid; within a preset period, calculate the difference between the global maximum and minimum values ​​of the normalized active power data as the peak-valley difference characteristic value to quantify the peak-valley fluctuation range of the load and identify the load characteristics of the power grid; S32: Photovoltaic power fluctuation pattern identification: Through power spectrum density analysis, the main fluctuation frequency components in photovoltaic power data are extracted, their energy proportions are calculated, and the periodic fluctuation patterns of output are identified based on their energy proportions. This is used to build a photovoltaic consumption constraint model and optimize the scheduling strategy. S33: Correlation analysis, using sliding windows as units, calculate the covariance of photovoltaic power generation and active power, and divide them by the product of their standard deviations to obtain the photovoltaic-load Pearson correlation coefficient; if the correlation coefficient is positive, it means that the photovoltaic output and load demand change synchronously; if it is negative, it means inverse fluctuation. According to the topological relationship, calculate the correlation coefficient between the active power of each node. Specifically, calculate the Pearson correlation coefficient between the active power of each node, process the active power data of each node through a sliding window method, calculate the covariance between each pair of nodes, and divide them by the product of their standard deviations to obtain the correlation coefficient, and generate an adjacency matrix through the correlation coefficient. Use the generated adjacency matrix to analyze the correlation distribution and connection pattern between the nodes. Through the matrix analysis method, identify the load propagation path in the distribution network, that is, determine the transfer of load from one node to other nodes, so as to identify the load propagation path and regional coupling characteristics, predict the load change trend and distribution to optimize the scheduling of distributed energy output and load demand.

[0025] Step S4 load forecast classification specifically includes the following steps: S41: Construct a load demand forecasting model, organize the pre-processed normalized active power data into sliding window data of historical periods according to time series, and form multi-dimensional time series input vector samples at fixed intervals. The samples contain active power values, node codes, and data quality flags with continuous time stamps, and are aligned into a space-time matrix according to the distribution network topology. A long-short-term memory network structure is constructed, consisting of an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is consistent with the historical time step, and it receives the active power sequence. The hidden layer selectively memorizes historical information and updates the cell state through a triple gating mechanism consisting of a forget gate, an input gate, and an output gate. The output layer is mapped to the load demand sequence for the future period, and the load demand sequence dimension matches the prediction duration. S42: Load demand forecasting and load state classification. The latest collected active power data is normalized and input into the trained load demand forecasting model. The historical information is filtered through the forgetting gate, the memory unit is updated by the input gate, and the hidden state is generated by the output gate to output the load demand sequence forecast result. The maximum value is extracted from the load demand forecast result as the predicted load peak, and the predicted load peak is divided by the rated total capacity of the distribution network and multiplied by 100% to obtain the predicted load rate in percentage form. The predicted load rate range is determined and the corresponding load state label is generated, including high load state, normal load state and low load state, and the load state label is bound to the load rate value. If the predicted load rate is ≥85%, it is a high load state. If it is 60%≤predicted load rate<85%, it is a normal load state. If the predicted load rate is <60%, it is a low load state. A recursive prediction strategy is used to balance the long-term prediction accuracy and error accumulation problems.

[0026] Step S5 load dynamic optimization classifies the load status in the previous step, generates high load, normal load and low load state label inputs, and drives the load dynamic optimization module. Specifically, it includes the following steps: S51: State-weight mapping: Construct a state-weight mapping table. If the system is in a high-load state, the objective function focuses on reducing network loss; if the system is in a normal-load state, the objective function focuses on balancing optimization cost and network loss; if the system is in a low-load state, the objective function focuses on cost. Receive the load status tag, search the status-weight mapping table in real time to obtain the preset weight value, and trigger a smooth weight transition if the status tags are the same for two consecutive cycles; S52: Adjust the particle swarm inertia weight, define the economic cost, network loss cost, and cost-network loss objective functions, and dynamically adjust the particle swarm inertia weight w according to the load state label; if it is in a high load state, accelerate the convergence speed to prioritize stability; if it is in a normal load state, set the balance between global search and local convergence capabilities; if it is in a low load state, set the enhanced global search capability to explore cost potential; The ratio of cost weight to network loss weight is used as the boundary condition for particle position update to limit the search space; S53: Construct a photovoltaic absorption constraint model based on a mixed integer programming model architecture. Define photovoltaic output, energy storage charging and discharging power, and main grid purchase power as continuous variables, and photovoltaic unit switching status and transformer tap position as integer variables. Convert the weight-driven load state-related photovoltaic absorption constraints into linear constraints. If the system is in a high-load state, the photovoltaic power generation power is limited to no more than 0.9 times the rated photovoltaic power generation power, and the energy storage is forced to discharge to make up for the shortfall; if the system is in a normal-load state, the photovoltaic output benchmark value is limited to 0.8 to 1.1 times the current load, and the energy storage charge and discharge power is limited to 50% of the rated value; if the system is in a low-load state, the photovoltaic power generation power is limited to no less than 0.95 times the load power, and energy storage is allowed to reversely supply power to the main grid.

[0027] Step S6: Energy efficiency scheduling analysis, specifically including the following steps: S61: Generates dispatch instructions, receives a PV consumption constraint model containing a load-state-driven objective function, continuous variables, and integer variables, converts the objective function and constraints into a mathematical format recognizable by the solver, calls the CPLEX solver, uses a branch-and-bound algorithm to traverse the feasible solution space, prioritizes integer variable combinations, calculates node currents and network losses, evaluates the objective function value, and selects a Pareto optimal solution set that satisfies the weight constraints. After the solution is completed, the specific values ​​of the continuous variables are extracted from the results returned by the solver. Based on the analyzed continuous variable values, specific dispatch instructions are generated. The dispatch instructions include: generating photovoltaic inverter power setting instructions, energy storage charging and discharging power instructions, main network power purchase plan and transformer gear adjustment instructions; each type of generated instruction is encapsulated into a standardized control frame according to node code, timestamp and execution priority, and the instructions are broadcast to photovoltaic power stations, energy storage controllers and load management terminals in a 1-second cycle. The deviation rate between the instructions and real-time data is checked. If the deviation exceeds 5%, the model is re-solved, and the dispatch strategy is synchronized with the energy-saving optimization dispatch system status of the multi-source distribution network; S62: Generate a difference heat map. Before the dispatch instruction is issued, collect the current and line resistance of each node in real time to calculate the node power loss. After the dispatch instruction is executed, calculate the node power loss and loss difference based on the optimized current value to generate a loss difference heat map. Use red areas to mark high-loss nodes with positive loss differences, and the color scale depth is proportional to the loss increment; use green areas to mark valid nodes with negative loss differences, and the color scale depth is proportional to the loss reduction; use gray areas to mark nodes with no obvious change in loss difference. Click a node in the loss difference heat map to display the scheduling instruction details of the node. It supports replaying the heat map changes before and after scheduling along the timeline to analyze the spatiotemporal effects of the strategy.

[0028] A multi-source distribution network energy-saving optimization and dispatching system based on big data processing, characterized by comprising a distribution network multi-source data acquisition module, a distribution network data preprocessing module, a load forecasting and classification module, a load dynamic optimization module and an energy efficiency dispatching and analysis module; The distribution network multi-source data acquisition module collects the active power data of each node in the distribution network and the photovoltaic power generation data of the photovoltaic power station; The load forecasting classification module builds a load demand forecasting model based on active power data, thereby predicting the load demand of the distribution network and classifying the load status by label; The load dynamic optimization module uses the adaptive weight coefficient mechanism in the improved particle swarm algorithm to optimize the load state analysis results after label classification, maps the load state classification results to cost weights and network loss weights, and constructs a photovoltaic absorption constraint model; The energy efficiency scheduling analysis module calls the CPLEX solver to solve the photovoltaic absorption constraint model, generates scheduling instructions, and generates a difference heat map based on the power loss of the distribution network before and after the scheduling instructions are taken effect.

[0029] A terminal device, characterized in that it includes a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the method according to any one of claims 1 to 7 is implemented.

[0030] A computer-readable storage medium, characterized in that a computer program is stored in the storage medium, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

[0031] The specific implementation method is as follows: S1: multi-source data collection of the distribution network, including distribution network active power collection and photovoltaic power generation power collection; the distribution network active power collection is connected to the three-phase power line through the distribution network node, and connected to the regional data concentrator through the power line carrier, and an independent device code is configured for the smart meter. It is mapped to the multi-source distribution network energy-saving optimization scheduling system database according to the node topology relationship, and the built-in current transformer of the smart meter is used to sense the line current in real time, and then calculate the active power; The photovoltaic power generation power collection is deployed on the DC side and AC side of the grid-connected inverter of the photovoltaic power station through a photovoltaic inverter monitoring device. The DC side uses a Hall sensor to measure the output current of the photovoltaic string, and the maximum power point is calculated in combination with the open-circuit voltage of the photovoltaic string to output the photovoltaic power generation power.

[0032] S2: Distribution network data preprocessing, including outlier detection and removal, data normalization, and noise filtering; specifically, the following steps: S21: Outlier detection and elimination, receiving data streams from smart meters and photovoltaic inverters, setting the sliding window width, calculating the mean and standard deviation of active power and photovoltaic power generation within the window point by point, and judging whether the current data point deviates from the mean by more than 3 times the standard deviation. If so, it is marked as an outlier and eliminated from the data stream; if not, it is marked as a normal value. In view of the fluctuation characteristics of photovoltaic power generation, an adaptive threshold is introduced in anomaly detection. Specifically, when processing distribution network data, a sliding window is set, the width of which is determined according to the time resolution and fluctuation characteristics of the data. The sliding window is used to calculate the mean and standard deviation of active power and photovoltaic power generation within the window point by point. Within the sliding window, the standard deviation of active power and photovoltaic power generation data is calculated respectively, and then the standard deviation within the sliding window is multiplied by a dynamic coefficient to obtain an adaptive threshold. The dynamic coefficient is adjusted according to the actual application scenario and data characteristics to enhance or weaken the role of the standard deviation in outlier detection. S22: Data normalization: Calculate the minimum and maximum values ​​of active power and photovoltaic power generation within a preset time period. Use a linear transformation formula to map the original data to the [0, 1] interval. Update the minimum and maximum values ​​every hour. S23: Noise filtering: For power data with obvious high-frequency noise interference, a sliding average algorithm with a window width of 5 seconds is used to calculate the mean of the data in the window as the output, and the maximum change slope of the data in the filtering window is limited.

[0033] S3: Data feature extraction, including active power fluctuation feature extraction, photovoltaic power generation power fluctuation pattern recognition and correlation analysis; specifically includes the following steps: S31: Extraction of active power fluctuation characteristics, including power fluctuation amplitude and peak-valley fluctuation range of load; taking a fixed time period as the statistical window, taking the difference between the maximum and minimum values ​​of the normalized active power data in the window, dividing it by the time period length, to obtain the power fluctuation amplitude per unit time; within a preset period, statistically calculating the difference between the global maximum and minimum values ​​of the normalized active power data as the peak-valley difference characteristic value to quantify the peak-valley fluctuation range of the load; the power fluctuation amplitude reflects the severity of the change in active power in a short period of time and is used to judge the stability of the power grid; the peak-valley fluctuation range quantifies the range of change of the load over a longer period and helps identify the load characteristics of the power grid. By analyzing these two characteristics, it is determined whether the power grid is in a stable operating state and whether there is an overload or underload risk; based on the power fluctuation amplitude and the load peak-valley fluctuation range, an optimized scheduling strategy is formulated. During periods with large power fluctuation amplitudes, the regulation capacity of the energy storage equipment needs to be increased or the generated power needs to be adjusted to smooth out the power fluctuations and ensure power grid stability. In areas with large peak-valley fluctuation ranges, demand-side management or flexible scheduling of distributed energy can be used to optimize load distribution, reduce peak-valley differences, and improve power grid operation efficiency; S32: Photovoltaic power generation power fluctuation pattern identification, through power spectrum density analysis, extract the main fluctuation frequency components in the photovoltaic power generation data, calculate their energy proportion, and identify the periodic fluctuation pattern, that is, the regular period of photovoltaic power generation over time, reflecting the power output law of photovoltaic power generation under the influence of natural factors; specifically, through power spectrum density analysis, extract the main fluctuation frequency components from the photovoltaic power generation data. Power spectrum density analysis can convert the time series data of photovoltaic power generation into frequency domain information, identify the power fluctuation intensity at different frequencies, and reflect the changing characteristics of photovoltaic power generation on different time scales. Then, the energy proportion of the extracted fluctuation frequency components is calculated. The energy proportion reflects the importance of each frequency component in the entire photovoltaic power generation power fluctuation, identifies the frequency component with the greatest impact on the photovoltaic power generation power fluctuation, determines the main periodic fluctuation pattern, and finally identifies the periodic fluctuation pattern of photovoltaic power based on the energy proportion. The subsequent periodic fluctuation pattern is used to construct a photovoltaic absorption constraint model and optimize the dispatching strategy to ensure that under different fluctuation modes, the power grid can efficiently absorb photovoltaic energy while maintaining stable operation.

[0034] S33: Correlation analysis. Using a sliding window as a unit, calculate the covariance of photovoltaic power generation and active power, and divide them by the product of their standard deviations to obtain the photovoltaic-load Pearson correlation coefficient. If the correlation coefficient is positive, it means that the photovoltaic output and load demand change synchronously. If it is negative, it means inverse fluctuation. Based on the topological relationship, calculate the Pearson correlation coefficient between the active power of each node. Process the active power data of each node in a sliding window manner, calculate the covariance between each pair of nodes, and divide them by the product of their standard deviations to obtain the correlation coefficient. Based on the calculated correlation coefficients between the active powers of each node, generate an adjacency matrix. The adjacency matrix is ​​used to represent the connection relationship and correlation strength between nodes. Each element in the matrix represents the correlation coefficient between the corresponding nodes. Use the generated adjacency matrix to analyze the correlation distribution and connection pattern between nodes. Through the matrix analysis method, identify the load propagation path in the distribution network, that is, determine the direction and intensity of load transmission from one node to other nodes. At the same time, analyze the coupling characteristics between regions and identify the interaction relationship between different regions.

[0035] S4: Load forecast classification, by building a load demand forecast model and using the load demand forecast model to predict and classify the load state; using the long short-term memory network model to predict the load demand, and classifying the load state based on the prediction results, generating high load, normal load, and low load state labels. These state labels are directly used as input to S5 to drive the load dynamic optimization module. Specifically, it includes the following steps: S41: Construct a load demand forecasting model, organize the pre-processed normalized active power data into sliding window data of historical periods according to time series, and form multi-dimensional time series input vector samples at fixed intervals. The samples contain active power values, node codes, and data quality flags with continuous time stamps, and are aligned into a space-time matrix according to the distribution network topology. A long-short-term memory network structure is constructed, consisting of an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is consistent with the historical time step, and it receives the active power sequence. The hidden layer selectively memorizes historical information and updates the cell state through a triple gating mechanism consisting of a forget gate, an input gate, and an output gate. The output layer is mapped to the load demand sequence for the future period, and the load demand sequence dimension matches the prediction duration. S42: Load demand forecasting and load state classification. The latest collected active power data is normalized and input into the trained load demand forecasting model. The historical information is filtered through the forgetting gate, the memory unit is updated by the input gate, and the hidden state is generated by the output gate to output the load demand sequence forecast result. The maximum value is extracted from the load demand forecast result as the predicted load peak, and the predicted load peak is divided by the rated total capacity of the distribution network and multiplied by 100% to obtain the predicted load rate in percentage form. The predicted load rate range is determined and the corresponding load state label is generated, including high load state, normal load state and low load state, and the load state label is bound to the load rate value. If the predicted load rate is ≥85%, it is a high load state. If it is 60%≤predicted load rate<85%, it is a normal load state. If the predicted load rate is <60%, it is a low load state. A recursive prediction strategy is used to balance the long-term prediction accuracy and error accumulation problems.

[0036] S5: Dynamic load optimization, through state-weight mapping, adjusting the particle swarm inertia weight, and building a photovoltaic absorption constraint model; according to different load states, the weight distribution of the objective function is determined through the state-weight mapping table, and then the inertia weight of the particle swarm algorithm is adjusted to build a photovoltaic absorption constraint model. This process ensures that the optimization strategy can be dynamically adjusted according to the real-time load state, thereby achieving multi-objective optimization of cost, network loss and photovoltaic absorption. The output of S4 provides a decision-making basis for S5, and the optimization results of S5 further improve the operating efficiency and energy saving effect of the distribution network. Specifically, it includes the following steps: S51: State-weight mapping: Construct a state-weight mapping table. If the system is in a high-load state, the objective function focuses on reducing network loss; if the system is in a normal-load state, the objective function focuses on balancing optimization cost and network loss; if the system is in a low-load state, the objective function focuses on cost. Receive the load status tag, search the status-weight mapping table in real time to obtain the preset weight value, and trigger a smooth weight transition if the status tags are the same for two consecutive cycles; S52: Adjust the particle swarm inertia weight, define the economic cost, network loss cost and cost-network loss objective function. The objective function mainly includes: economic cost and network loss cost. The economic cost mainly considers the operating cost of the power system, including the cost of purchasing electricity from the main grid, the charging and discharging cost of energy storage equipment, etc. The network loss cost is the loss cost generated by the power grid in the process of transmitting and distributing electricity, which is usually proportional to the square of the current and the line resistance. The particle swarm inertia weight w is dynamically adjusted according to the load state label. Under high load conditions, the objective function focuses on reducing network loss. It is necessary to accelerate the convergence speed of the particle swarm algorithm to prioritize the stability of the power grid. At this time, the inertia weight is set to a smaller value to reduce the inertia of the particles and make them converge to the local optimal solution faster. Under normal load conditions, the objective function needs to balance the economic cost and network loss cost. Therefore, the particle swarm algorithm needs to balance the global search ability and local convergence ability. At this time, the inertia weight is set to a medium value to take into account both global search and local search. Under low load conditions, the objective function focuses on optimizing the economic cost. Therefore, the global search ability of the particle swarm algorithm needs to be enhanced to tap the cost potential. At this time, the inertia weight is set to a larger value to increase the inertia of the particles, so that it can explore the solution space more widely. The ratio of cost weight to network loss weight is used as the boundary condition for particle position update to limit the search space; S53: Construct a photovoltaic absorption constraint model based on a mixed integer programming model architecture. Define photovoltaic output, energy storage charging and discharging power, and main grid purchase power as continuous variables, and photovoltaic unit switching status and transformer tap position as integer variables. Convert the weight-driven load state-related photovoltaic absorption constraints into linear constraints. Based on the load state (high load, normal load, low load), combined with a weight mapping table, the load state-related requirements are converted into specific constraints. The constraints reflect the different requirements for photovoltaic absorption under different load states. Under high load conditions, the photovoltaic power generation power is limited to no more than a certain proportion of the rated power, and energy storage is forced to discharge to make up for the shortfall. Under normal load conditions, a photovoltaic output reference value range and energy storage charge and discharge power limits are set. Under low load conditions, the photovoltaic power generation power is limited to no less than a certain proportion of the load power, and energy storage is allowed to reversely supply power to the main grid. Specifically, if the load is high, the photovoltaic power generation power is limited to no more than 0.9 times the rated photovoltaic power generation power, and energy storage is forced to discharge to make up for the shortfall. If the load is normal, the photovoltaic output reference value is limited to 0.8 to 1.1 times the current load, and the energy storage charge and discharge power is limited to 50% of the rated value. If the load is low, the photovoltaic power generation power is limited to no less than 0.95 times the load power, and energy storage is allowed to reversely supply power to the main grid.

[0037] S6: Energy efficiency scheduling analysis, generating scheduling instructions and generating difference heat maps. Specifically, it includes the following steps: S61: Generates dispatch instructions, receives a photovoltaic absorption constraint model including an objective function driven by the load state, continuous variables, and integer variables, and converts the objective function and constraints into a mathematical format recognizable by the solver; the objective function is a multi-objective optimization function driven by the load state, including economic cost and network loss cost. The economic cost involves the main grid power purchase cost and the energy storage charging and discharging cost. The network loss cost is related to the square of the current at each node and the line resistance. The photovoltaic absorption constraint model is based on a mixed integer programming architecture and includes continuous variables and integer variables. The CPLEX solver is called, and a branch-and-bound algorithm is used to traverse the feasible solution space, giving priority to integer variable combinations, calculating the current and network loss at each node, and evaluating the objective function value. The branch-and-bound algorithm is used to systematically search all possible solutions to ensure that the optimal solution that meets all constraints is found. Finally, a Pareto optimal solution set that meets the weight constraints is selected from all feasible solutions; After the solution is completed, the specific values ​​of the continuous variables are extracted from the results returned by the solver. Based on the continuous variable values ​​obtained by the analysis, specific scheduling instructions are generated. The scheduling instructions include: photovoltaic inverter power setting instructions, which are used to adjust the output power of the photovoltaic power station, energy storage charging and discharging power instructions, which are used to control the charging and discharging operations of the energy storage equipment, main grid power purchase plan, which is used to arrange the amount of electricity purchased from the main grid, and transformer gear adjustment instructions, which are used to optimize the operating status of the transformer; the various types of instructions generated are encapsulated into standardized control frames according to node coding, timestamp, and execution priority, and the instructions are broadcast to the photovoltaic power station, energy storage controller, and load management terminal in a period of 1 second. The deviation rate between the instructions and real-time data is checked. If the deviation exceeds 5%, the model is re-solved, and the scheduling strategy is synchronized with the energy-saving optimization scheduling system status of the multi-source distribution network; S62: Generate a difference heat map. Before the dispatch instruction is issued, collect the current and line resistance of each node in real time to calculate the node power loss. After the dispatch instruction is executed, calculate the node power loss and loss difference based on the optimized current value to generate a loss difference heat map. Use red areas to mark high-loss nodes with positive loss differences, and the color scale depth is proportional to the loss increment; use green areas to mark valid nodes with negative loss differences, and the color scale depth is proportional to the loss reduction; use gray areas to mark nodes with no obvious change in loss difference. Click a node in the loss difference heat map to display the scheduling instruction details of the node. It supports replaying the heat map changes before and after scheduling along the timeline to analyze the spatiotemporal effects of the strategy.

Claims

1. A multi-source distribution network energy-saving optimization scheduling method based on big data processing, characterized in that: The following steps are involved: S1: Multi-source data collection of distribution network, including active power collection of distribution network and photovoltaic power generation power collection; S2: Distribution network data preprocessing, including outlier detection and removal, data normalization, and noise filtering; S3: Data feature extraction, including active power fluctuation feature extraction, photovoltaic power generation power fluctuation pattern recognition and correlation analysis; S4: Load forecast classification, by building a load demand forecast model and classifying the load state through the load demand forecast model prediction; S5: Dynamic load optimization, through state-weight mapping, adjusting particle swarm inertia weight, and building a photovoltaic absorption constraint model; S6: Energy efficiency scheduling analysis, generating scheduling instructions and generating difference heat maps.

2. The multi-source distribution network energy-saving optimization scheduling method based on big data processing according to claim 1 is characterized in that: Step S1: collecting multi-source data of the distribution network, including collecting active power of the distribution network and collecting photovoltaic power generation power; The distribution network active power collection is connected to the three-phase power line through the distribution network node, and connected to the regional data concentrator through the power line carrier. The smart meter is configured with an independent device code, mapped to the multi-source distribution network energy-saving optimization scheduling system database according to the node topology relationship, and the line current is sensed in real time by the built-in current transformer of the smart meter to calculate the active power. The photovoltaic power generation power collection is deployed on the DC side and AC side of the grid-connected inverter of the photovoltaic power station through a photovoltaic inverter monitoring device. The DC side uses a Hall sensor to measure the output current of the photovoltaic string, and the maximum power point is calculated in combination with the open-circuit voltage of the photovoltaic string to output the photovoltaic power generation power.

3. The multi-source distribution network energy-saving optimization scheduling method based on big data processing according to claim 1 is characterized in that: Step S2: pre-processing the distribution network data, specifically including the following steps: S21: Outlier detection and elimination: Receive data streams from smart meters and photovoltaic inverters, set the sliding window width, calculate the mean and standard deviation of active power and photovoltaic power generation within the window point by point, and determine whether the current data point deviates from the mean by more than 3 times the standard deviation. If so, it is marked as an outlier and eliminated from the data stream; if not, it is marked as a normal value. In view of the fluctuation characteristics of photovoltaic power generation, an adaptive threshold is introduced during anomaly detection. Specifically, a sliding window is set, the width of which is determined according to the time resolution and fluctuation characteristics of the data. The sliding window is used to calculate the mean and standard deviation of active power and photovoltaic power generation within the window point by point. Within the sliding window, the standard deviation of the active power and photovoltaic power generation data is calculated respectively. Then, the standard deviation within the sliding window is multiplied by a dynamic coefficient to obtain the adaptive threshold. S22: Data normalization: Calculate the minimum and maximum values ​​of active power and photovoltaic power generation within a preset time period. Use a linear transformation formula to map the original data to the [0, 1] interval. Update the minimum and maximum values ​​every hour. S23: Noise filtering: For power data with obvious high-frequency noise interference, a sliding average algorithm with a window width of 5 seconds is used to calculate the mean of the data in the window as the output, and the maximum change slope of the data in the filtering window is limited.

4. The multi-source distribution network energy-saving optimization scheduling method based on big data processing according to claim 1 is characterized in that: Step S3 data feature extraction specifically includes the following steps: S31: Extract active power fluctuation characteristics, including power fluctuation amplitude and load peak-valley fluctuation range; using a fixed time period as a statistical window, take the difference between the maximum and minimum values ​​of the normalized active power data within the window, divide it by the time period length, and obtain the power fluctuation amplitude per unit time, which is used to judge the stability of the power grid; within a preset period, calculate the difference between the global maximum and minimum values ​​of the normalized active power data as the peak-valley difference characteristic value to quantify the peak-valley fluctuation range of the load and identify the load characteristics of the power grid; S32: Photovoltaic power fluctuation pattern identification: Through power spectrum density analysis, the main fluctuation frequency components in photovoltaic power data are extracted, their energy proportions are calculated, and the periodic fluctuation patterns of output are identified based on their energy proportions. This is used to build a photovoltaic consumption constraint model and optimize the scheduling strategy. S33: Correlation analysis, using sliding windows as units, calculate the covariance of photovoltaic power generation and active power, and divide them by the product of their standard deviations to obtain the photovoltaic-load Pearson correlation coefficient; if the correlation coefficient is positive, it means that the photovoltaic output and load demand change synchronously; if it is negative, it means inverse fluctuation. According to the topological relationship, calculate the correlation coefficient between the active power of each node. Specifically, calculate the Pearson correlation coefficient between the active power of each node, process the active power data of each node through a sliding window method, calculate the covariance between each pair of nodes, and divide them by the product of their standard deviations to obtain the correlation coefficient, and generate an adjacency matrix through the correlation coefficient. Use the generated adjacency matrix to analyze the correlation distribution and connection pattern between the nodes. Through the matrix analysis method, identify the load propagation path in the distribution network, that is, determine the transfer of load from one node to other nodes, so as to identify the load propagation path and regional coupling characteristics, predict the load change trend and distribution to optimize the scheduling of distributed energy output and load demand.

5. The multi-source distribution network energy-saving optimization scheduling method based on big data processing according to claim 1 is characterized in that: Step S4 load forecast classification specifically includes the following steps: S41: Construct a load demand forecasting model, organize the pre-processed normalized active power data into sliding window data of historical periods according to time series, and form multi-dimensional time series input vector samples at fixed intervals. The samples contain active power values, node codes, and data quality flags with continuous time stamps, and are aligned into a space-time matrix according to the distribution network topology. A long-short-term memory network structure is constructed, consisting of an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is consistent with the historical time step, and it receives the active power sequence. The hidden layer selectively memorizes historical information and updates the cell state through a triple gating mechanism consisting of a forget gate, an input gate, and an output gate. The output layer is mapped to the load demand sequence for the future period, and the load demand sequence dimension matches the prediction duration. S42: Load demand forecasting and load state classification. The latest collected active power data is normalized and input into the trained load demand forecasting model. The historical information is filtered through the forgetting gate, the memory unit is updated by the input gate, and the hidden state is generated by the output gate to output the load demand sequence forecast result. The maximum value is extracted from the load demand forecast result as the predicted load peak, and the predicted load peak is divided by the rated total capacity of the distribution network and multiplied by 100% to obtain the predicted load rate in percentage form. The predicted load rate range is determined and the corresponding load state label is generated, including high load state, normal load state and low load state, and the load state label is bound to the load rate value. If the predicted load rate is ≥85%, it is a high load state. If it is 60%≤predicted load rate<85%, it is a normal load state. If the predicted load rate is <60%, it is a low load state. A recursive prediction strategy is used to balance the long-term prediction accuracy and error accumulation problems.

6. The multi-source distribution network energy-saving optimization scheduling method based on big data processing according to claim 1 is characterized in that: Step S5 load dynamic optimization classifies the load status in the previous step, generates high load, normal load and low load state label inputs, and drives the load dynamic optimization module. Specifically, it includes the following steps: S51: State-weight mapping: Construct a state-weight mapping table. If the system is in a high-load state, the objective function focuses on reducing network loss; if the system is in a normal-load state, the objective function focuses on balancing optimization cost and network loss; if the system is in a low-load state, the objective function focuses on cost. Receive the load status tag, search the status-weight mapping table in real time to obtain the preset weight value, and trigger a smooth weight transition if the status tags are the same for two consecutive cycles; S52: Adjust the particle swarm inertia weight, define the economic cost, network loss cost, and cost-network loss objective functions, and dynamically adjust the particle swarm inertia weight w according to the load state label; if it is in a high load state, accelerate the convergence speed to prioritize stability; if it is in a normal load state, set the balance between global search and local convergence capabilities; if it is in a low load state, set the enhanced global search capability to explore cost potential; The ratio of cost weight to network loss weight is used as the boundary condition for particle position update to limit the search space; S53: Construct a photovoltaic absorption constraint model based on a mixed integer programming model architecture. Define photovoltaic output, energy storage charging and discharging power, and main grid purchase power as continuous variables, and photovoltaic unit switching status and transformer tap position as integer variables. Convert the weight-driven load state-related photovoltaic absorption constraints into linear constraints. If the system is in a high-load state, the photovoltaic power generation power is limited to no more than 0.9 times the rated photovoltaic power generation power, and the energy storage is forced to discharge to make up for the shortfall; if the system is in a normal-load state, the photovoltaic output benchmark value is limited to 0.8 to 1.1 times the current load, and the energy storage charge and discharge power is limited to 50% of the rated value; if the system is in a low-load state, the photovoltaic power generation power is limited to no less than 0.95 times the load power, and energy storage is allowed to reversely supply power to the main grid.

7. The multi-source distribution network energy-saving optimization scheduling method based on big data processing according to claim 1 is characterized in that: Step S6: Energy efficiency scheduling analysis, specifically including the following steps: S61: Generates dispatch instructions, receives a PV consumption constraint model containing a load-state-driven objective function, continuous variables, and integer variables, converts the objective function and constraints into a mathematical format recognizable by the solver, calls the CPLEX solver, uses a branch-and-bound algorithm to traverse the feasible solution space, prioritizes integer variable combinations, calculates node currents and network losses, evaluates the objective function value, and selects a Pareto optimal solution set that satisfies the weight constraints. After the solution is completed, the specific values ​​of the continuous variables are extracted from the results returned by the solver. Based on the analyzed continuous variable values, specific dispatch instructions are generated. The dispatch instructions include: generating photovoltaic inverter power setting instructions, energy storage charging and discharging power instructions, main network power purchase plan and transformer gear adjustment instructions; each type of generated instruction is encapsulated into a standardized control frame according to node code, timestamp and execution priority, and the instructions are broadcast to photovoltaic power stations, energy storage controllers and load management terminals in a 1-second cycle. The deviation rate between the instructions and real-time data is checked. If the deviation exceeds 5%, the model is re-solved, and the dispatch strategy is synchronized with the energy-saving optimization dispatch system status of the multi-source distribution network; S62: Generate a difference heat map. Before the dispatch instruction is issued, collect the current and line resistance of each node in real time to calculate the node power loss. After the dispatch instruction is executed, calculate the node power loss and loss difference based on the optimized current value to generate a loss difference heat map. Use red areas to mark high-loss nodes with positive loss differences, and the color scale depth is proportional to the loss increment; use green areas to mark valid nodes with negative loss differences, and the color scale depth is proportional to the loss reduction; use gray areas to mark nodes with no obvious change in loss difference. Click a node in the loss difference heat map to display the scheduling instruction details of the node. It supports replaying the heat map changes before and after scheduling along the timeline to analyze the spatiotemporal effects of the strategy.

8. A multi-source distribution network energy-saving optimization and dispatching system based on big data processing, characterized by: It includes distribution network multi-source data acquisition module, distribution network data preprocessing module, load forecasting and classification module, load dynamic optimization module and energy efficiency scheduling and analysis module; The distribution network multi-source data acquisition module collects the active power data of each node in the distribution network and the photovoltaic power generation data of the photovoltaic power station; The load forecasting classification module builds a load demand forecasting model based on active power data, thereby predicting the load demand of the distribution network and classifying the load status by label; The load dynamic optimization module uses the adaptive weight coefficient mechanism in the improved particle swarm algorithm to optimize the load state analysis results after label classification, maps the load state classification results to cost weights and network loss weights, and constructs a photovoltaic absorption constraint model; The energy efficiency scheduling analysis module calls the CPLEX solver to solve the photovoltaic absorption constraint model, generates scheduling instructions, and generates a difference heat map based on the power loss of the distribution network before and after the scheduling instructions are taken effect.

9. A terminal device, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images