Power grid electricity price intelligent prediction method based on multi-model fusion algorithm

Abstract

The invention discloses a power grid electricity price intelligent prediction method based on a multi-model fusion algorithm, and belongs to the technical field of power grid management, and the method specifically comprises the steps: dividing a target prediction region into a plurality of non-overlapping grid units according to the geographic position, and setting a unique code for each unit; historical electricity price, real-time load and power transmission state data are collected, and a space electricity price weight reflecting the electricity price fluctuation influence degree of adjacent areas is generated; fusing the space electricity price weight and the time sequence electricity price feature to obtain a space-time fusion electricity price feature vector; by means of a pre-trained dynamic prediction model, outputting an electricity price prediction value in a future target period according to the feature vector of each grid unit; once the deviation between the predicted value of a grid unit and the actual electricity price of an adjacent unit exceeds a set threshold value, a space weight dynamic correction mechanism is triggered, the weight is recalculated, and a prediction result is iteratively updated; the accuracy of power grid electricity price prediction is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid management, and particularly to an intelligent power grid electricity price prediction method based on a multi-model fusion algorithm. Background Art

[0002] With the gradual opening of the power market and the large-scale grid connection of new energy, the volatility of power grid electricity prices has increased significantly. Electricity price prediction is the core link of power market operation, directly affecting the bidding strategies of power generation enterprises, the economy of power grid dispatching, and the electricity costs of end-users. Traditional electricity price prediction methods mainly rely on the statistical analysis of historical electricity price data. However, with the increase in the complexity of the power system, especially the introduction of new elements such as distributed energy and demand response, the accuracy and robustness of a single prediction model are difficult to meet the actual needs.

[0003] Currently, the mainstream methods in the field of power grid electricity price prediction can be divided into three categories: methods based on statistical models, such as ARIMA and GARCH. These methods are suitable for stationary time series but have limited ability to capture electricity price mutations and non-linear relationships; methods based on machine learning, such as support vector machines and random forests, which improve the prediction effect through feature engineering but are difficult to handle spatio-temporal coupling features; methods based on deep learning, such as LSTM and Transformer, which can learn long-term dependence relationships but have insufficient adaptability to the dynamic changes of real-time data.

[0004] Most existing models independently process electricity price data in different regions, ignoring the dynamic impact of the power grid topology structure and the direction of power transmission on electricity price fluctuations; and once the fusion model is trained, its weight allocation rule is fixed and cannot automatically adjust the prediction logic according to real-time electricity price deviations. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent power grid electricity price prediction method based on a multi-model fusion algorithm to solve the following technical problems: Existing models ignore the dynamic impact of the power grid topology structure and the direction of power transmission on electricity price fluctuations and cannot automatically adjust the prediction logic according to real-time electricity price deviations.

[0006] The purpose of the present invention can be achieved through the following technical solutions: An intelligent power grid electricity price prediction method based on a multi-model fusion algorithm includes the following steps: According to the transmission line topology diagram of the target prediction area, divide the target prediction area into multiple non-overlapping grid units, and set a unique geographical location code in each grid unit; Collect historical electricity price data, real-time load data of each grid cell, and power transmission status data of adjacent grid cells to generate the spatial electricity price weight of each grid cell, where the spatial electricity price weight represents the degree of influence of any grid cell on the electricity price fluctuations of adjacent grid cells; Extract the corresponding time series electricity price features according to the historical electricity price data of each grid cell, and fuse the time series electricity price features of each grid cell with the spatial electricity price weight to generate a spatio-temporal fused electricity price feature vector; For each grid cell, according to its spatio-temporal fused electricity price feature vector, output the electricity price prediction value for the future target period through a pre-trained dynamic prediction model; When the deviation between the predicted value of any grid cell and the actual electricity price of adjacent grid cells exceeds the set threshold, trigger the spatial weight dynamic correction mechanism, recalculate the spatial electricity price weight of the grid cell and iteratively update the prediction result.

[0007] As a further solution of the present invention: The method for dividing the grid cells is as follows: Obtain the transmission line topology relationship diagram in the power grid dispatching system, identify all line intersection points as initial boundary nodes, and generate polygon boundaries centered on each intersection point along the extension direction of the transmission lines. Adjacent polygons share at least one side; Statistically analyze the average hourly load data of each sub-region within a set duration in the target area, calculate the maximum load value per unit area as the load density benchmark, divide the load density into multiple levels according to the percentile of the benchmark, and the area of each grid cell is inversely proportional to its load density. The higher the load density, the smaller the grid area; when the power grid is expanded or the line is transformed, the grid boundary is dynamically adjusted according to the updated topology relationship.

[0008] As a further solution of the present invention: The generation process of the spatial electricity price weight is as follows: Select the hourly electricity price sequences of the target grid and each adjacent grid within the past set number of days, calculate the matching degree of their fluctuation trends, calculate the ratio of the number of days when the adjacent grid synchronously rises to the set number of days when the target grid's electricity price rises, and mark it as the correlation. If it is determined to be a strong correlation, the higher the correlation, the higher the initial weight; Real-time monitor the power flow data in the power grid dispatching system. If the target grid currently outputs electric energy to a certain adjacent grid, multiply the initial weight of the adjacent grid by a set attenuation coefficient. If the target grid is currently an electric energy input party, the weight of the adjacent grid remains unchanged; the final spatial electricity price weight is the weighted average after correlation correction and transmission direction correction, and the sum of the weights of all adjacent grid cells is normalized.

[0009] As a further solution of the present invention: The method for extracting the time series electricity price features is as follows: Obtain the historical electricity price data of each grid cell, split the electricity price data of each day within a set past duration by hour, and calculate the average value of the electricity price within the same hour as the 24-hour benchmark period curve; then calculate the amplitude of the deviation of the electricity price from the benchmark period curve at the same hour on weekdays and holidays respectively. If the average increase of the electricity price on holidays exceeds the first set ratio, it is marked as a significant difference feature. Using the benchmark period curve as the baseline, subtract the baseline value from the weekday features, and add the second set ratio of the baseline value as a correction amount to the holiday features, and finally splice them in chronological order by hour to form a time dimension feature vector.

[0010] As a further solution of the present invention: the generation process of the spatio-temporal fusion electricity price feature vector includes: Standardize the spatial electricity price weights of each grid so that their value ranges from 0 to 1, and then map the weight values to the influence coefficients on a logarithmic scale. Multiply each element in the time dimension feature vector by the influence coefficient of the corresponding grid to obtain the eigenvalue, calculate the sum of the eigenvalues of all grids at a certain time point, and scale the eigenvalues of each grid proportionally so that the sum of the eigenvalues is equal to a preset global benchmark value.

[0011] As a further solution of the present invention: the triggering condition of the spatial weight dynamic correction mechanism is: Obtain the actual electricity price data of the adjacent grids of any target grid in the current period. When the predicted electricity price of the target grid is higher than the average value of the actual electricity prices of the adjacent grids and the difference exceeds the first threshold, it is determined as an abnormally high price area, and then the correction is triggered. When the predicted electricity price of the target grid is lower than the minimum value of the actual electricity prices of the adjacent grids and the difference exceeds the second threshold, it is determined as an abnormally low price area, and then the correction is triggered.

[0012] As a further solution of the present invention: the spatial weight dynamic correction mechanism includes: Freeze the prediction result of the current target grid, re-collect the latest power transmission status data, re-calculate the spatial electricity price weights based on the updated transmission direction, and input the new weights into the dynamic prediction model to generate a corrected electricity price prediction value.

[0013] As a further solution of the present invention: the training process of the dynamic prediction model includes: Extract the spatio-temporal feature vectors of each grid from the historical data, and associate their predicted values before correction, actual values after correction, and the reason labels for triggering correction to form an annotated data set. Randomly tamper with the spatial weight values of some samples to simulate weight distortion caused by incorrect transmission direction or invalid correlation calculation. After the model outputs the predicted value, compare it with the true value. If the deviation exceeds the threshold, reverse-adjust the model parameters; learn the prediction deviation pattern caused by abnormal spatial weights and automatically compensate for the weight calculation error when outputting the predicted value.

[0014] As a further solution of the present invention: store the timestamp, grid code, deviation type, and correction amplitude of each correction event in the database, and aggregate them according to the geographical location code to generate a heat map; obtain the fault alarm records of the power grid dispatching system within the same time period. If the occurrence time of the correction event in a certain grid coincides with the alarm time such as line overload and transformer failure, it is marked as a strongly associated event; count the number of times of triggering corrections in the target grid within a week. If it exceeds the set number and there is no associated alarm record, send a check request to the operation and maintenance system.

[0015] The beneficial effects of the present invention: By dividing the target prediction area into grid units with unique geographical location codes and innovatively introducing spatial electricity price weights to characterize the dynamic influence relationship between grid units, the present invention effectively solves the problem of insufficient spatial correlation modeling in the prior art; by real-time collecting the power transmission status data of adjacent grid units and calculating the fluctuation trend matching degree, and dynamically correcting the weights in combination with the power transmission direction, it overcomes the defect that the weight allocation in traditional multi-model fusion is fixed and cannot adapt to the real-time changes of the power grid; by fusing the spatial electricity price weights with the time series electricity price characteristics to generate a spatio-temporal fusion electricity price feature vector and designing a dynamic correction mechanism for spatial weights including the determination of high-price anomalies and low-price anomalies, it realizes a rapid response to emergencies such as holidays and extreme weather, significantly improving the prediction accuracy; further, by constructing a training sample set with correction labels and optimizing the dynamic prediction model by means of adversarial training, the model has the ability to automatically compensate for weight calculation errors, and at the same time generates a spatial weight anomaly event map by associating correction events with power grid fault alarm data, providing an accurate positioning basis for potential abnormal areas for power grid operation and maintenance, thereby greatly improving the robustness and practicality of the system while ensuring the real-time nature of prediction. Brief Description of the Drawings

[0016] The following further describes the present invention with reference to the drawings.

[0017] Figure 1 It is a flow schematic diagram of the present invention. Detailed Embodiments

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] Please refer to Figure 1 As shown, the present invention is an intelligent power grid electricity price prediction method based on a multi-model fusion algorithm, including the following steps: Based on the power grid GIS system, obtain the topological structure data of the target prediction area. Taking the intersection points of substations and transmission lines as reference nodes, use the Voronoi diagram algorithm to generate multiple non-overlapping grid cells. Each grid cell is assigned a geographical location code with a hierarchical structure, and the coding format is "area code - voltage level - grid serial number". For example, "BJ-110kV-005" represents the 5th grid cell with a voltage level of 110kV in the Beijing area.

[0020] In the data collection stage, the system obtains the historical electricity price data of each grid cell through the power market trading platform, with a sampling frequency of 15 minutes per time; collects real-time load data through the SCADA system, including parameters such as active power and reactive power; at the same time, accesses the power transmission status data of the EMS system to record the power flow direction and transmission power between adjacent grid cells. After preprocessing these data, an improved grey relational analysis method is used to calculate the spatial electricity price weight. The specific process is as follows: Select the electricity price time series of the target grid cell and adjacent grid cells in the recent 30 days, calculate their grey relational coefficients, and then correct them in combination with the real-time power flow direction. If the target grid is an electricity output party, apply a decay factor of 0.6 to the corresponding weight, and finally obtain a standardized spatial electricity price weight within the range of 0-1.

[0021] The spatio-temporal feature fusion link is realized by using the feature cross technology: The time series electricity price feature extraction uses a multi-scale analysis method to decompose the historical electricity price data into daily, weekly, and seasonal cycles, and obtain a 24-dimensional periodic feature vector; After converting the spatial electricity price weight into a non-linear influence coefficient through the sigmoid function, perform a Hadamard product operation with the time feature vector to generate a 72-dimensional spatio-temporal fusion feature vector. This feature vector fully retains the spatio-temporal correlation characteristics of electricity price fluctuations.

[0022] The dynamic prediction model adopts a cascaded network structure. The first layer is a bidirectional GRU network to capture time dependencies, the second layer is a graph attention network to process spatial associations, and the output layer introduces an adaptive weighting mechanism. During model training, a curriculum learning strategy is adopted. First, pre-training is performed using historical data, and then incremental learning is used for updating. During the prediction execution phase, the system monitors the prediction results of each grid cell in real time. When it is detected that the deviation between the predicted electricity price of a certain grid cell and the actual electricity price of adjacent cells exceeds 15% for three consecutive cycles, a correction mechanism is triggered: the prediction process of this grid is paused, the latest power grid state data is re-collected, the spatial electricity price weights are updated based on the sliding window method, and the corrected features are input into the prediction model to generate new results. This process is executed cyclically until the deviation drops below 10%. The system also records the detailed information of each correction event for online optimization of model parameters and power grid anomaly diagnosis.

[0023] In a preferred embodiment of the present invention, the method for dividing grid cells is as follows: Obtain the transmission line topology diagram from the power grid dispatching system. Use image recognition algorithms to analyze this diagram and accurately identify all line intersection points. These intersection points are determined as initial boundary nodes, which are the basis for constructing grid cells.

[0024] Taking each identified intersection point as the center, according to the actual extension direction of the transmission line, use the polygon generation algorithm to generate polygon boundaries. During the generation process, strictly follow the rule that adjacent polygons share at least one side, so as to ensure the spatial integrity and continuity of the grid cell division and provide a reliable regional division basis for subsequent power-related analysis.

[0025] For the target area, count the average hourly load data of each sub-area within a set time period. The set time period can be flexibly determined according to the operating characteristics of the power grid and analysis requirements, such as selecting the past month or quarter. Process the large amount of collected load data, calculate the maximum load value per unit area, and use it as the load density benchmark.

[0026] Divide the load density into multiple levels according to the percentile of the load density benchmark. On this basis, determine the area of each grid cell. Set an inverse relationship between the grid cell area and the load density, that is, the higher the load density area, the smaller the area of the divided grid cells. For example, in the core business district of a city with large and frequent load demands, the corresponding grid area is small; while in remote suburban areas with relatively stable power consumption demands and low load densities, the grid area is large.

[0027] When the power grid is expanded or the line is reconstructed, the updated topological relationship data is obtained in time. The specially designed algorithm is used to analyze and process these new data, so as to dynamically adjust the grid boundary and ensure that the grid unit division always matches the real-time structure and operation status of the power grid, providing accurate basic data support for subsequent electricity price forecasting and power grid management.

[0028] In another preferred embodiment of the present invention, the process of generating the spatial electricity price weight is: Calculate the matching degree and correlation of fluctuation trends: select the hourly electricity price sequence of the target grid and each adjacent grid within the past set number of days. The set number of days is usually determined based on the stability of the power market and the periodicity of electricity price fluctuations, such as 30 days or 60 days. Use data analysis algorithms to deeply analyze these electricity price sequences and calculate the matching degree of fluctuation trends between the target grid and the adjacent grids. At the same time, calculate the ratio of the number of days when the electricity price of the target grid rises synchronously to the set number of days, and mark the ratio as correlation. When the ratio is high, it is judged as a strong correlation. The higher the correlation, the higher the initial weight given to the adjacent grid.

[0029] Correct weights based on power flow: Establish real-time data connection with the power grid dispatching system to continuously monitor power flow data. If the target grid is currently exporting power to an adjacent grid, the initial weight of the adjacent grid is multiplied by a preset attenuation coefficient based on the possible losses in the power transmission process and the impact on the electricity price. If the target grid is currently importing power, the weight of the adjacent grid remains unchanged.

[0030] Calculate the final spatial electricity price weight and normalize it: perform weighted average calculation on the weights after correlation correction and transmission direction correction to obtain the spatial electricity price weight of each adjacent grid unit relative to the target grid. Finally, normalize the sum of the weights of all adjacent grid units so that the sum of all weights is 1 to ensure the rationality and comparability of the weights, so as to accurately quantify the spatial correlation between electricity prices.

[0031] In another preferred embodiment of the present invention, the method for extracting the time series electricity price characteristics is: Collect historical electricity price data for each grid unit. Select a set period of time in the past, such as the past year, and divide the daily electricity price data in this period by hour. For each hour, summarize all the electricity price data corresponding to that hour in previous years and calculate its average value. For example, calculate the average electricity price at 1:00 on January 1, 1:00 on February 1, and so on, and 1:00 on December 31 every year, and calculate the average value for each hour in this way, and finally form a complete 24-hour benchmark cycle curve. This curve reflects the average electricity price level of the grid unit at different times of the day under normal circumstances.

[0032] Further distinguish between electricity prices on weekdays and holidays. Calculate the deviation of the electricity price at the same hour on weekdays and holidays from the previously obtained benchmark cycle curve. The specific calculation method is to take the electricity price at 10 o'clock on a certain holiday as an example, subtract the electricity price corresponding to 10 points in the benchmark cycle curve from the electricity price, and then divide it by the electricity price at 10 points in the benchmark cycle curve to obtain the deviation ratio of the electricity price at 10 o'clock on the holiday relative to the benchmark. Sum up and average the deviation ratios of the same hour on all holidays. If this average increase exceeds the first set ratio (such as 15%), the hour is marked as having significant difference characteristics. This helps to identify periods when holiday electricity prices differ greatly from regular working day electricity prices.

[0033] Taking the benchmark cycle curve as the reference baseline, the electricity price characteristics of weekdays and holidays are processed. When processing the weekday characteristics, the baseline value (i.e., the electricity price of the corresponding hour in the benchmark cycle curve) is deducted from the electricity price data corresponding to each hour to highlight the changes in the weekdays relative to the normal average level. When processing holiday characteristics, based on the electricity price data corresponding to each hour, the second set ratio of the baseline value (such as 10%) is added as a correction amount to reflect the adjustment direction and magnitude of holiday electricity prices relative to normal conditions. Finally, these processed data are spliced ​​together in order of hours to form a complete time dimension feature vector, which fully reflects the change characteristics of electricity prices at different times relative to the benchmark level under different date types.

[0034] In a preferred case of this embodiment, the process of generating the spatiotemporal fusion electricity price feature vector includes: First, the spatial electricity price weight of each grid is obtained. Since the weight values ​​from different sources may have dimensional differences, these weights are standardized. Through a specific standardization formula, such as the maximum-minimum standardization method, the value range of the weight value is uniformly adjusted to between 0 and 1. Afterwards, in order to better reflect the degree of influence of the weight on the electricity price characteristics, the standardized weight value is mapped to the influence coefficient on the logarithmic scale. The specific mapping method can be to convert through a logarithmic function, for example, using a logarithmic function with a base of 10 to operate on the weight value, and converting the linear change of the weight value into a coefficient value that can better reflect the relative influence on the logarithmic scale.

[0035] Multiply each element in the time - dimension feature vector by the influence coefficient after logarithmic mapping of the corresponding grid, so as to obtain the eigenvalue of each grid at a certain time point. For example, multiply the 5th element in the time - dimension feature vector (corresponding to the 5th hour) by the influence coefficient of grid No. 3 to get the eigenvalue of grid No. 3 at the 5th hour. Calculate the sum of the eigenvalues of all grids at a certain time point. According to the preset rules, scale the eigenvalue of each grid proportionally so that the sum of the eigenvalues of all grids at this time point is equal to the preset global reference value (such as 100). Through such a scaling operation, it is ensured that the eigenvalues between different grids are comparable and overall conform to the preset global quantization standard, and finally generate a fused electricity price feature vector that can comprehensively reflect spatio - temporal characteristics, providing effective input data for the subsequent electricity price prediction model.

[0036] In another preferred embodiment of the present invention, the trigger condition of the spatial weight dynamic correction mechanism is: Real - time obtain the actual electricity price data of the adjacent grids of any target grid in the current period. This process is docked with the data interface of the power market trading platform to ensure timely and accurate acquisition of the latest electricity price information. The range of adjacent grids is determined according to the power grid topology structure, covering the surrounding grids directly connected to the target grid through transmission lines.

[0037] Calculate the average value of the actual electricity prices of the adjacent grids. Compare the predicted electricity price of the target grid with this average value. If the predicted electricity price of the target grid is higher than the average value of the actual electricity prices of the adjacent grids and the difference between the two exceeds the first threshold (for example, the first threshold is set to 0.1 yuan / kWh), it is determined that the area where the target grid is located is an abnormally high - price area, and at this time, the spatial weight dynamic correction mechanism is triggered.

[0038] Find the minimum value of the actual electricity prices of the adjacent grids. When the predicted electricity price of the target grid is lower than this minimum value and the difference between the two exceeds the second threshold (such as the second threshold is set to 0.08 yuan / kWh), it is determined that the area where the target grid is located is an abnormally low - price area, and the spatial weight dynamic correction mechanism is also triggered.

[0039] In another preferred embodiment of the present invention, the spatial weight dynamic correction mechanism includes: Immediately freeze the prediction result of the current target grid to prevent subsequent operations based on possibly inaccurate prediction data. This freezing operation ensures that during the correction process, no more problems will be caused due to the continuous influence of the original incorrect prediction result.

[0040] Re - collect the latest power transmission status data. By establishing a real - time connection with the power grid dispatching system, obtain the latest information including the power flow direction and transmission power of the transmission lines. These data are crucial for accurately calculating the spatial electricity price weights.

[0041] Based on the updated data such as the power transmission direction, a specific algorithm is used to recalculate the spatial electricity price weights. For example, according to the new transmission direction, combined with the topological structure of the power grid and the historical electricity price correlation data, the influence weights of each adjacent grid on the electricity price of the target grid are re-evaluated. The newly calculated weights are input into the dynamic prediction model, and the model recalculates the electricity price of the target grid based on the new weights to generate a corrected electricity price prediction value.

[0042] In a preferred case of this embodiment, the training process of the dynamic prediction model includes: Extract the spatio-temporal feature vectors of each grid from the historical data. The spatio-temporal feature vectors include the electricity price change characteristics in different periods (such as daily cycle, weekly cycle, seasonal cycle) in the time dimension, and the association characteristics with adjacent grids in the space dimension. Associate these feature vectors with their predicted values before correction, the actual values after correction, and the reason labels for triggering the correction. For example, the reason labels may include information such as "the electricity price fluctuation is triggered by a power generation failure in the adjacent area". Organize these data into a labeled data set to provide rich samples for model training.

[0043] During the training process, randomly tamper with the spatial weight values of some samples. Simulate the weight distortion situation that may occur in actual operation due to incorrect transmission direction or invalid correlation calculation. The model processes these samples with abnormal weights and outputs predicted values, and compares the predicted values with the true values (i.e., the actual values after correction). If the deviation exceeds a pre-set threshold (such as the relative error between the predicted electricity price and the true electricity price exceeds 5%), the model parameters are adjusted backward according to the error backpropagation algorithm. Through multiple such training processes, the model can learn the prediction deviation patterns caused by abnormal spatial weights.

[0044] After repeated training, the model is enabled to automatically compensate for the weight calculation error when outputting the predicted value. When encountering an actual abnormal spatial weight situation, the model can self-correct the prediction result according to the learned deviation pattern, improving the accuracy and stability of the prediction.

[0045] It should be noted that the present invention stores information such as the timestamp, grid code, deviation type (abnormal high price or abnormal low price), and correction amplitude of each correction event in the database. The database adopts an efficient storage structure to ensure the rapid storage and query of data. For example, a relational database is used, and corresponding indexes are established to speed up data retrieval.

[0046] Aggregate the corrected event data according to the geographical location code to generate a heat map. The heat map intuitively shows the occurrence frequency and severity of corrected events in different regions. For example, the darker the color, the more frequent the corrected events occur or the greater the correction amplitude in that region. Through the heat map, the operation and maintenance personnel can quickly locate the regions where the electricity price fluctuations are relatively concentrated in the power grid.

[0047] Obtain the fault alarm records of the power grid dispatching system during the same time period. Compare the occurrence time of the corrected event with the alarm times such as line overload and transformer failure. If the occurrence time of the corrected event in a certain grid coincides with the alarm time, it is marked as a strongly correlated event. This helps to analyze the potential connection between abnormal electricity price fluctuations and physical faults in the power grid.

[0048] Count the number of times the target grid triggers corrections within a week. If it exceeds the set number of times (such as set to 5 times) and there is no associated alarm record, send a check request to the operation and maintenance system to conduct an in-depth investigation of the grid and related transmission lines and equipment to determine whether there are potential power system problems causing abnormal electricity price fluctuations.

[0049] The above has described an embodiment of the present invention in detail, but the content described is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of the present invention application should still fall within the patent coverage scope of the present invention.

Claims

1. A method for intelligent prediction of power grid electricity prices based on a multi-model fusion algorithm, characterized in that: The following steps are involved: According to the transmission line topology diagram of the target prediction area, the target prediction area is divided into multiple non-overlapping grid units, and a unique geographic location code is set in each grid unit; Collect historical electricity price data, real-time load data and power transmission status data of adjacent grid units for each grid unit, and generate spatial electricity price weights for each grid unit, wherein the spatial electricity price weights represent the degree to which any grid unit is affected by electricity price fluctuations of adjacent grid units; According to the historical electricity price data of each grid unit, the corresponding time series electricity price characteristics are extracted, and the time series electricity price characteristics of each grid unit are fused with the spatial electricity price weight to generate a spatiotemporal fusion electricity price feature vector; For each grid unit, according to its spatiotemporal fusion electricity price feature vector, the pre-trained dynamic prediction model is used to output the electricity price forecast value for the future target period; When the deviation between the predicted value of any grid unit and the actual electricity price of the adjacent grid unit exceeds the set threshold, the spatial weight dynamic correction mechanism is triggered, the spatial electricity price weight of the grid unit is recalculated and the prediction result is iteratively updated.

2. The method for intelligent prediction of power grid electricity prices based on a multi-model fusion algorithm according to claim 1 is characterized in that: The grid unit division method is: Obtain the topological relationship diagram of the transmission lines in the power grid dispatching system, identify all line intersections as initial boundary nodes, and generate polygonal boundaries along the extension direction of the transmission lines with each intersection as the center, with adjacent polygons sharing at least one edge; Statistics are collected for each sub-area within the set time period of the target area, and the maximum load value per unit area is calculated as the load density benchmark. The load density is divided into multiple levels according to the percentile of the benchmark. The area of ​​each grid unit is inversely proportional to its load density. The higher the load density, the smaller the grid area. When the power grid is expanded or the line is transformed, the grid boundary is dynamically adjusted according to the updated topological relationship.

3. The method for intelligent prediction of power grid electricity prices based on multi-model fusion algorithm according to claim 1 is characterized in that: The generation process of the spatial electricity price weight is: Select the hourly electricity price sequence of the target grid and each adjacent grid within the past set number of days, calculate the matching degree of the fluctuation trends of the two, calculate the ratio of the number of days when the electricity price of the target grid rises synchronously with the number of set days, mark it as correlation, and judge it as strong correlation. The higher the correlation, the higher the initial weight. The power flow data of the power grid dispatching system is monitored in real time. If the target grid currently outputs power to an adjacent grid, the initial weight of the adjacent grid is multiplied by the set attenuation coefficient. If the target network is currently the power input party, the weight of the adjacent grid remains unchanged. The final spatial electricity price weight is the weighted average after correlation correction and transmission direction correction, and the sum of the weights of all adjacent grid units is normalized.

4. The method for intelligent prediction of power grid electricity prices based on a multi-model fusion algorithm according to claim 1 is characterized in that: The method for extracting the time series electricity price characteristics is: Obtain the historical electricity price data of each grid unit, divide the daily electricity price data within the past set time period by hour, calculate the average electricity price within the same hour as the 24-hour benchmark cycle curve; then calculate the deviation of the electricity price of the same hour on weekdays and holidays from the benchmark cycle curve. If the average increase in holiday electricity prices exceeds the first set ratio, it is marked as a significant difference feature; Taking the benchmark cycle curve as the baseline, the baseline value is deducted from the weekday characteristics, and the second set ratio of the baseline value is added to the holiday characteristics as a correction amount, and finally spliced ​​into a time dimension feature vector in hourly order.

5. The method for intelligent prediction of power grid electricity prices based on multi-model fusion algorithm according to claim 4 is characterized in that: The generation process of the spatiotemporal fusion electricity price feature vector includes: The spatial electricity price weight of each grid is standardized so that its value range is between 0 and 1, and then the weight value is mapped to the impact coefficient on a logarithmic scale; Multiply each element in the time dimension feature vector by the influence coefficient of the corresponding grid to obtain the eigenvalue, calculate the sum of the eigenvalues ​​of all grids at a certain time point, and scale the eigenvalue of each grid proportionally so that the sum of the eigenvalues ​​is equal to the preset global benchmark value.

6. The method for intelligent prediction of power grid electricity prices based on multi-model fusion algorithm according to claim 1 is characterized in that: The triggering conditions of the spatial weight dynamic correction mechanism are: Obtain the actual electricity price data of the adjacent grids of any target grid in the current period; When the predicted electricity price of the target grid is higher than the average of the actual electricity prices of the adjacent grids and the difference exceeds the first threshold, it is determined to be an abnormally high price area, and a correction is triggered; When the predicted electricity price of the target grid is lower than the lowest value of the actual electricity price of the adjacent grids and the difference exceeds the second threshold, it is determined to be an abnormally low price area, and correction is triggered.

7. The method for intelligent prediction of power grid electricity prices based on multi-model fusion algorithm according to claim 1 is characterized in that: The dynamic modification mechanism of spatial weights includes: Freeze the prediction results of the current target grid, re-collect the latest power transmission status data, recalculate the spatial electricity price weights based on the updated transmission direction, and input the new weights into the dynamic prediction model to generate the revised electricity price prediction value.

8. The method for intelligent prediction of power grid electricity prices based on multi-model fusion algorithm according to claim 7 is characterized in that: The training process of the dynamic prediction model includes: Extract the spatiotemporal feature vector of each grid from the historical data, and associate its predicted value before correction, actual value after correction, and the label of the cause that triggered the correction to form a labeled data set; The spatial weight values ​​of some samples are randomly tampered with to simulate the weight distortion caused by incorrect transmission direction or failed correlation calculation. The model is required to output the predicted value and compare it with the true value. If the deviation exceeds the threshold, the model parameters are adjusted in the opposite direction. The prediction deviation pattern caused by spatial weight anomalies is learned, and the weight calculation error is automatically compensated when the predicted value is output.

9. The method for intelligent prediction of power grid electricity prices based on multi-model fusion algorithm according to claim 8 is characterized in that: The timestamp, grid code, deviation type, and correction amplitude of each correction event are stored in the database, and aggregating them according to the geographic location code to generate a heat map; the fault alarm records of the power grid dispatching system in the same time period are obtained. If the correction event time of a grid coincides with the alarm time of line overload, transformer failure, etc., it is marked as a strongly correlated event; The number of times the target grid triggers corrections within a week is counted. If it exceeds the set number and there is no associated alarm record, a check request is sent to the operation and maintenance system.

Images