A power transaction supply and demand prediction method and system based on a regional meteorological large model

Abstract

The application discloses a power transaction supply and demand prediction method and system based on a regional meteorological large model, relates to the technical field of power transaction prediction, and comprises the following steps: acquiring historical power load, new energy power generation and meteorological observation data of a target region, inputting a pre-trained regional meteorological large model, performing feature cross calculation on the three types of data through a multi-source data fusion layer, and generating an hourly gridded basic meteorological element field; calling an energy response mapping module, respectively mapping each meteorological grid into an air conditioner load response curve, a wind power output response curve and a hydropower output correction coefficient, time-aligning the mapping results to generate an energy supply side and a consumption side prediction data set, and generating a supply and demand gap time sequence according to the difference between the two. The method solves the prediction deviation and time sequence misplacement problems of the prior art, realizes deep fusion of multi-source data and accurate mapping of meteorological elements, improves the accuracy of power transaction supply and demand prediction, and adapts to the demand of power transaction dispatching.

Description

Technical Field

[0001] This invention belongs to the field of power trading forecasting technology, specifically a power trading supply and demand forecasting method and system based on a regional meteorological large model. Background Technology

[0002] Electricity trading supply and demand forecasting is an important support for the orderly operation of the electricity market. In the existing technology, electricity trading supply and demand forecasting is mostly carried out by combining meteorological data and electricity-related data. Typically, traditional meteorological forecasting models are used to obtain meteorological data, and then the meteorological data is simply superimposed or superficially correlated with historical electricity load and new energy power generation data to achieve supply and demand forecasting.

[0003] Existing technical solutions suffer from insufficient multi-source data fusion. They lack in-depth feature cross-processing of historical power load data, historical renewable energy generation data, and historical meteorological observation data, failing to form accurate basic data reflecting the relationship between regional meteorology and energy. Furthermore, the meteorological data has low temporal resolution and is not converted into a rasterized distribution format, making it difficult to accurately match the spatiotemporal variations in energy output and load. In addition, existing technologies do not establish dedicated mapping mechanisms for the correspondence between different meteorological elements and energy supply and demand, nor do they perform time-series alignment of various energy-related forecast results, leading to biased forecast data and inaccurate calculations of the supply-demand gap.

[0004] It is necessary to achieve deep fusion of multi-source data to generate high temporal resolution rasterized meteorological element data, establish accurate mapping relationship between meteorological elements and energy output and load, and solve the problems of weak data correlation, inconsistent time series and insufficient accuracy in existing forecasts. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a power trading supply and demand forecasting method based on a regional meteorological large-scale model, comprising: Historical power load data, historical renewable energy generation data, and historical meteorological observation data of the target area are acquired and input into a pre-trained regional meteorological big model. The multi-source data fusion layer in the regional meteorological big model performs feature cross-calculation on the historical power load data, historical renewable energy generation data, and historical meteorological observation data to generate a basic meteorological element field of the target area. The basic meteorological element field includes a temperature distribution grid, a relative humidity distribution grid, and a surface wind speed distribution grid with a time resolution of hour. Based on the basic meteorological element field, the energy response mapping module in the regional meteorological big model is called to map the hourly temperature distribution grid as an air conditioning load response curve, the hourly surface wind speed distribution grid as a wind power output response curve, and the hourly relative humidity distribution grid as a hydropower output correction coefficient. The air conditioning load response curve, wind power output response curve and hydropower output correction coefficient are time-series aligned to generate the energy supply-side output prediction dataset and the energy consumption-side load prediction dataset for the target area. Based on the difference between the energy supply-side output forecast dataset and the energy consumption-side load forecast dataset, a time series of the power trading supply and demand gap in the target region is generated.

[0006] Furthermore, the data is input into a pre-trained regional meteorological big data model. Through the multi-source data fusion layer in the regional meteorological big data model, feature cross-calculation is performed on the historical power load data, historical new energy power generation data, and historical meteorological observation data to generate the basic meteorological element field of the target area, including: The historical power load data includes power consumption curves for various industries; the historical new energy power generation data includes irradiance conversion curves for photovoltaic power plants and wind speed-power curves for wind turbines; and the historical meteorological observation data includes spatiotemporal distribution records of temperature, humidity, air pressure, and cloud cover. The historical meteorological observation data were subjected to quality control processing to remove sensor drift data and missing period data, and to complete the time series of missing period data. The historical power load data is divided into industrial load subsequence, commercial load subsequence and residential load subsequence according to industry category, and the historical new energy power generation data is divided into photovoltaic power output subsequence and wind power output subsequence. The industrial load subsequence, commercial load subsequence, and residential load subsequence are correlated with the historical meteorological observation data with a time resolution of hours to extract the meteorological sensitive feature vector on the load side. The photovoltaic power output subsequence and the wind power output subsequence are coupled with the historical meteorological observation data with a time resolution of hours for physical mechanism analysis to extract the meteorological driving feature vector on the power supply side. The meteorological sensitive feature vector from the load side and the meteorological driving feature vector from the power supply side are input into the multi-source data fusion layer of the regional meteorological big model. The data is then processed by a multilayer perceptron network to perform nonlinear transformation, and the output is the basic meteorological element field containing the spatial distribution characteristics of air temperature, relative humidity, and surface wind speed.

[0007] Furthermore, mapping the temperature distribution grid with a time resolution of hours to an air conditioning load response curve includes: The average temperature of the residential area and the average temperature of the commercial area are extracted from the temperature distribution grid with a time resolution of hours. Based on the average temperature of the residential area and the preset temperature comfort range, the cooling demand index and heating demand index of the residential area are calculated. The cooling demand index is the cumulative value of the difference between the residential area temperature and the upper limit of comfort, and the heating demand index is the cumulative value of the difference between the residential area temperature and the lower limit of comfort. Based on the average temperature of the commercial area and the preset commercial temperature sensitivity coefficient, the air conditioning load adjustment of the commercial area is calculated. The air conditioning load adjustment of the commercial area is equal to the deviation between the temperature of the commercial area and the median of the comfort range multiplied by the commercial temperature sensitivity coefficient. The total air conditioning load forecast value of the target area is generated by weighted summing of the cooling demand index and heating demand index of the residential area and the air conditioning load adjustment of the commercial area. The total air conditioning load forecast values ​​are connected in hourly order to form the air conditioning load response curve.

[0008] Furthermore, mapping the surface wind speed distribution grid with a time resolution of hours to a wind power output response curve includes: Spatial downscaling is performed on the surface wind speed distribution grid with a time resolution of hour to convert the low-resolution grid data into a high-resolution wind speed sequence corresponding to the wind farm turbine location coordinates. For each wind speed value in the high-resolution wind speed sequence, a preset wind turbine power curve lookup table is called to convert the wind speed value into the theoretical output value of a single wind turbine. The wind turbine power curve lookup table contains the corresponding power relationship between the cut-in wind speed, the rated wind speed, and the cut-out wind speed. The theoretical output values ​​of all wind turbines in a single wind farm are summed to obtain the total theoretical output value of the single wind farm. The total theoretical output values ​​of all wind farms in the target area are then summarized to obtain the total theoretical wind power output value of the target area. Based on the turbulence intensity data of the surface wind speed distribution grid extracted from the basic meteorological element field, the total theoretical wind power output is reduced by multiplying the total theoretical wind power output by a correction factor that is inversely proportional to the turbulence intensity. The reduced total theoretical wind power output is connected in hourly order to form the wind power output response curve.

[0009] Furthermore, mapping the relative humidity distribution grid with a time resolution of hours to a hydropower output correction coefficient includes: Obtain the watershed hydrological forecast data of hydropower stations within the target area, wherein the watershed hydrological forecast data includes the predicted inflow and water level of upstream reservoirs; By fusing and analyzing the relative humidity distribution grid with a time resolution of hours with the watershed hydrological forecast data, the high relative humidity areas that trigger heavy precipitation and their duration can be identified. Based on the coverage and duration of the high relative humidity area, the predicted inflow of the upstream reservoir is corrected to generate a corrected predicted inflow. Substituting the corrected inflow forecast and reservoir water level forecast into the power output calculation formula of the hydropower station, the corrected power output value considering the impact of precipitation is calculated. The ratio of the corrected output value to the baseline output value (which does not consider the impact of precipitation) is the result of the hydropower output correction coefficient.

[0010] Furthermore, the air conditioning load response curve, wind power output response curve, and hydropower output correction coefficient are time-series aligned to generate an energy supply-side output prediction dataset and an energy consumption-side load prediction dataset for the target area, including: The air conditioning load response curve, wind power output response curve, and hydropower output correction coefficient are mapped to the same time coordinate system, with the smallest unit of the time coordinate system being the hour. The hydropower output correction coefficient is applied to the preset hydropower reference output curve to generate the corrected hydropower output curve. The hydropower reference output curve is the output curve of the hydropower station under the condition of no precipitation according to the reservoir scheduling plan. The modified hydropower output curve is added to the wind power output response curve, and the baseline output prediction values ​​of thermal power and other power sources are added to generate the energy supply-side output prediction dataset. The air conditioning load response curve is added to the predicted values ​​of other rigid loads besides the air conditioning load, and the data is summarized to generate the energy consumption side load prediction dataset. The generated energy supply-side output forecast dataset and energy consumption-side load forecast dataset are respectively subjected to smoothing filtering to remove abnormal fluctuation points introduced by data acquisition noise.

[0011] Further, based on the difference calculation results between the energy supply-side power output forecast dataset and the energy consumption-side load forecast dataset, a time series of the power trading supply-demand gap in the target region is generated, including: Subtract the corresponding hourly data point from the energy supply-side output forecast dataset to obtain the net power surplus or deficit value for each hour. A threshold for determining supply and demand balance is set. When the absolute value of the net power surplus or deficit is less than the threshold, the value of the corresponding hour is set to zero, which is regarded as a state of supply and demand balance. The net power surplus or deficit value after zeroing is accumulated and integrated to generate a cumulative supply and demand gap curve, which reflects the cumulative deviation of power supply and demand over a period of time. The net power surplus or deficit value is encoded together with the cumulative supply and demand deficit curve to generate the power trading supply and demand deficit time series containing instantaneous deficit information and cumulative deficit information.

[0012] Furthermore, it also includes a rolling revision step for supply and demand forecasts based on real-time weather updates: The system accesses the short-term and nowcast data output by the regional meteorological big model, which includes hourly updated meteorological element forecasts for the next 24 hours. The short-term and nowcast data is input into the rolling forecast interface of the regional meteorological big model to replace the original medium- and long-term meteorological forecast data. The incremental learning mechanism in the regional meteorological big data model is triggered, and the model weights are fine-tuned using the latest real-time power load data and real-time new energy power generation data to adapt to the latest meteorological change trends. The fine-tuned regional meteorological model is invoked to re-execute the forecasting process from generating the basic meteorological element field to generating the time series of electricity trading supply and demand gap; The regenerated time series of power trading supply and demand gap is compared with the previous forecast results using differential analysis, and the revised incremental data of supply and demand forecast is output.

[0013] Furthermore, the step of fine-tuning the model weights using the latest real-time power load data and real-time renewable energy generation data to adapt to the latest meteorological trends includes: Calculate the deviation rate between the latest real-time power load data and the predicted load value at the previous moment, and the deviation rate between the latest real-time renewable energy generation data and the predicted power generation value at the previous moment; The deviation rate is compared with a preset alarm threshold to filter out abnormal data points whose deviation rate exceeds the alarm threshold. Extract the spatiotemporal features corresponding to the abnormal data points and input them as negative samples into the loss function of the regional meteorological big model; By adjusting the weights of the fully connected layers related to the mapping of meteorological elements in the regional meteorological big model through the backpropagation algorithm, the confidence of the model in the meteorological characteristics that lead to the abnormal data points is reduced. The process of calculating the repetition rate, filtering outlier data points, inputting negative samples, and adjusting the weights of the fully connected layers continues until the repetition rate converges to within the alarm threshold, thus completing the fine-tuning of the model.

[0014] Furthermore, the present invention also includes a power trading supply and demand forecasting system based on a regional meteorological big data model. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the power trading supply and demand forecasting method based on a regional meteorological big data model as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Historical power load data, historical renewable energy generation data, and historical meteorological observation data of the target area are input into a pre-trained regional meteorological model. The model's multi-source data fusion layer performs feature cross-calculation on these three types of data, generating a basic meteorological element field containing hourly-level temperature distribution grids, relative humidity distribution grids, and surface wind speed distribution grids. This feature cross-calculation achieves deep fusion of multi-source data, rather than simple overlay, fully exploring the inherent correlations between the three types of data. The hourly-level rasterized distribution accurately captures subtle meteorological changes in different time periods and regions of the target area, enabling meteorological data to more accurately match the spatiotemporal variations of energy supply and demand. This solves the problems of insufficient multi-source data fusion, inadequate meteorological data resolution, and lack of rasterization in existing technologies, which result in weak correlations between meteorological and energy data.

[0016] Based on the energy response mapping module of the regional meteorological big data model, hourly temperature distribution grids are mapped to air conditioning load response curves, hourly surface wind speed distribution grids are mapped to wind power output response curves, and hourly relative humidity distribution grids are mapped to hydropower output correction coefficients. These three types of results are then time-series aligned to generate energy supply-side output forecast datasets and energy consumption-side load forecast datasets. The time series of the power trading supply-demand gap is calculated based on the difference between the two datasets. Establishing dedicated energy response mapping relationships for different meteorological elements makes the forecasting of various energy outputs and loads more targeted, avoiding the forecasting bias caused by the single mapping mode in existing technologies. Time-series alignment eliminates the misalignment of various forecast data in the time dimension, keeping the forecast data for the energy supply side and consumption side synchronized. This makes the generation of the supply-demand gap time series more closely resemble the actual scenario, solving the problem of insufficient gap calculation accuracy caused by time-series misalignment in existing technologies. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of a power trading supply and demand forecasting method based on a regional meteorological large model as described in this invention. Figure 2 A flowchart for mapping air conditioning load response curves; Figure 3 A flowchart generated for a time-series alignment and supply-demand forecasting dataset; Figure 4 A time-series comparison chart of electricity supply and demand; Figure 5 A comparison chart showing the effect of rolling correction for load forecasting. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 The system acquires historical power load data, historical renewable energy generation data, and historical meteorological observation data for the target area. This data is then input into a pre-trained regional meteorological model. Within the model, a multi-source data fusion layer performs feature cross-calculation on the input historical power load data, renewable energy generation data, and meteorological observation data. This calculation process generates a basic meteorological element field describing the meteorological conditions of the target area. This basic meteorological element field has an hourly temporal resolution and includes temperature distribution, relative humidity distribution, and surface wind speed distribution expressed in raster form. Based on this, the energy response mapping module built into the regional meteorological model is invoked. This module maps the hourly temperature distribution raster to an air conditioning load response curve reflecting temperature-sensitive load changes, maps the hourly surface wind speed distribution raster to a wind power output response curve reflecting wind energy resource fluctuations, and maps the hourly relative humidity distribution raster to a hydropower output correction coefficient used to correct hydropower generation forecasts. The generated air conditioning load response curve, wind power output response curve, and hydropower output correction coefficient are time-series aligned to integrate and generate a forecast dataset of energy supply-side output and energy consumption-side load for the target region over a future period. By calculating the difference between the energy supply-side output forecast dataset and the energy consumption-side load forecast dataset at each time point, a time series reflecting the power supply-demand gap in electricity trading, reflecting power surplus or shortage, is obtained, providing a quantitative basis for electricity market trading decisions.

[0020] In one embodiment of the present invention, historical power load data includes power consumption curves for various industries within the target area, historical renewable energy generation data includes irradiance conversion curves for photovoltaic power plants and wind speed-power curves for wind turbine generators, and historical meteorological observation data covers the spatiotemporal distribution records of elements such as temperature, humidity, air pressure, and cloud cover. When processing historical meteorological observation data, a quality control process is executed to remove drift data caused by sensor malfunctions and to identify and mark periods with missing data. For the marked missing periods, time series interpolation methods are used to complete the corresponding meteorological data, forming a continuous and complete historical meteorological observation data time series. Historical power load data is broken down into industrial load subsequences, commercial load subsequences, and residential load subsequences according to industry classification standards.

[0021] Historical renewable energy power generation data is broken down into photovoltaic (PV) output subsequences and wind power output subsequences based on power generation type. The industrial load subsequence, commercial load subsequence, and residential load subsequence are subjected to correlation analysis with historical meteorological observation data (with hourly time resolution) after completion processing. Correlation coefficients between different meteorological factors and various load sequences are calculated to extract load-side meteorological sensitivity feature vectors characterizing the load's sensitivity to meteorological conditions. The PV output subsequence and wind power output subsequence are subjected to physical mechanism coupling analysis with historical meteorological observation data of the same time resolution. For example, irradiance sequences are correlated with PV output, and wind speed sequences are correlated with wind power output to extract meteorological driving feature vectors that drive changes in renewable energy power generation. The load-side meteorological sensitivity feature vectors and the power-side meteorological driving feature vectors are input in parallel to the multi-source data fusion layer of the regional meteorological big data model. Within the multi-source data fusion layer, a multilayer perceptron network receives these feature vectors, learns and fuses meteorological correlation features from the load side and the power supply side through its internal multilayer nonlinear transformation processing, and finally outputs a basic meteorological element field containing the spatial distribution features of temperature, relative humidity and surface wind speed.

[0022] In the specific implementation, a provincial power grid in East China was selected as the target area. Historical power load data included industrial, commercial, and residential power consumption curves for 96 points (one point every 15 minutes) across the province from July 1st to August 31st, 2023. Historical renewable energy generation data included irradiance conversion curves of 20 centralized photovoltaic power stations and wind speed-power curves of 1500 grid-connected wind turbines during the same period. Historical meteorological observation data included hourly spatiotemporal distribution records of temperature, humidity, air pressure, and cloud cover collected by 120 automatic weather stations in the region during the same period. Quality control processing was performed on the historical meteorological observation data. Based on the physically reasonable threshold range set by the sensors, sensor drift data where temperature data was consistently above 50 degrees Celsius or below -10 degrees Celsius due to equipment failure were identified and removed. Simultaneously, periods with continuous data loss exceeding 3 hours due to communication interruptions were marked. For the marked missing periods, a linear interpolation method was used to fill in the missing data using data from two valid moments before and after the missing point, forming a continuous and complete historical meteorological observation data time series. Historical electricity load data is broken down according to the national economic industry classification standard into an industrial load subsequence represented by steel and chemical industries, a commercial load subsequence represented by shopping malls and office buildings, and a residential load subsequence represented by urban and rural residents' electricity consumption. Historical renewable energy power generation data is broken down according to power generation type into a photovoltaic power output subsequence reflecting solar energy conversion and a wind power output subsequence reflecting wind energy conversion.

[0023] Correlation analysis was performed on the industrial, commercial, and residential load subsequences with historical meteorological observation data (with hourly time resolution) after completion processing. Pearson correlation coefficients were calculated between each load subsequence and the corresponding temperature and humidity data, and a meteorological sensitivity feature vector on the load side, composed of these correlation coefficients, was extracted. Physical mechanism coupling analysis was performed between the photovoltaic output subsequence and the total radiation value in historical meteorological observation data, and between the wind power output subsequence and the 10-meter-high wind speed in historical meteorological observation data, extracting meteorological driving feature vectors on the power supply side characterizing photovoltaic conversion efficiency and wind turbine wind energy utilization rate. In some embodiments, the correlation analysis used Pearson correlation coefficients for calculation. The correlation coefficient between the industrial load subsequence and temperature data was 0.65, the commercial load subsequence and temperature data was 0.72, and the residential load subsequence and temperature data was 0.80. The correlation coefficients between the industrial load subsequence, commercial load subsequence, and residential load subsequence and humidity data were 0.30, 0.25, and 0.40, respectively. The load-side meteorological sensitivity feature vector is a six-dimensional vector, whose elements are, in order, the correlation coefficients between industrial load and temperature, commercial load and temperature, residential load and temperature, industrial load and humidity, commercial load and humidity, and residential load and humidity. Optionally, a physical mechanism coupling analysis is used to establish a mapping relationship between the irradiance conversion curve of the photovoltaic power station and the total radiation value in meteorological observation data, with a conversion coefficient of 0.18. A mapping relationship is also established between the wind turbine's wind speed-power curve and the wind speed at a height of 10 meters in meteorological observation data, with a conversion coefficient of 0.85. The power supply-side meteorological drive feature vector is a two-dimensional vector, whose elements are the photovoltaic irradiance conversion coefficient and the wind turbine wind speed conversion coefficient.

[0024] It can be understood that the meteorological sensitive feature vector from the load side and the meteorological driving feature vector from the power supply side are input in parallel to the multi-source data fusion layer of the regional meteorological large-scale model. The multi-source data fusion layer consists of a multilayer perceptron network. In specific implementation, the multilayer perceptron network includes one input layer, two hidden layers, and one output layer. The input layer has 8 neurons, corresponding to the 6 dimensions of the load-side meteorological sensitive feature vector and the 2 dimensions of the power supply-side meteorological driving feature vector. The first hidden layer contains 64 neurons, the second hidden layer contains 32 neurons, and the number of neurons in the output layer is determined by the number of spatial grid points in the basic meteorological element field. The nonlinear transformation processing of the multilayer perceptron network is implemented through activation functions; the hidden layer uses the ReLU activation function, and the output layer uses a linear activation function. The forward propagation formula of the multilayer perceptron network is: ; ; ; Wherein: symbol X represents the input feature vector, symbol W1 represents the weight matrix from the input layer to the first hidden layer, symbol b1 represents the bias vector of the first hidden layer, symbol H1 represents the output of the first hidden layer, symbol W2 represents the weight matrix from the first hidden layer to the second hidden layer, symbol b2 represents the bias vector of the second hidden layer, symbol H2 represents the output of the second hidden layer, symbol W3 represents the weight matrix from the second hidden layer to the output layer, symbol b3 represents the bias vector of the output layer, and symbol Y represents the basic meteorological element field data generated by the output layer. Optionally, the training of the multilayer perceptron network uses mean squared error as the loss function, and minimizes the difference between the predicted meteorological field and the actual observed meteorological field through the backpropagation algorithm and the Adam optimizer.

[0025] In one embodiment of the present invention, see [reference] Figure 2 From the generated hourly temperature distribution raster, average temperatures representing residential and commercial areas are extracted based on Geographic Information System (GIS) information. A predefined temperature comfort range with clearly defined upper and lower limits is established. The average temperature of the residential area is compared with this comfort range, and the cumulative temperature difference between the residential area and the upper limit is calculated as the cooling demand index; simultaneously, the cumulative temperature difference between the residential area and the lower limit is calculated as the heating demand index. A predefined commercial temperature sensitivity coefficient is established, reflecting the load response intensity of commercial activities to temperature changes. The average temperature of the commercial area is compared with the median of the comfort range, and the deviation is calculated. This deviation is then multiplied by the commercial temperature sensitivity coefficient to obtain the air conditioning load adjustment amount for the commercial area. The calculated cooling demand index, heating demand index, and air conditioning load adjustment amount for the residential and commercial areas are weighted and summed according to a predefined weight ratio to generate the total air conditioning load forecast for the target area in a specific hour. The total air conditioning load forecast values ​​calculated hourly are connected in chronological order to form a complete air conditioning load response curve.

[0026] Spatial downscaling is performed on the hourly-level surface wind speed distribution raster. An interpolation algorithm converts the relatively low-resolution raster data into a high-resolution wind speed time series corresponding to the specific wind turbine coordinates of each wind farm within the target area. For each hourly wind speed value in the high-resolution wind speed series, a pre-defined wind turbine power curve lookup table is consulted. This lookup table defines the mapping relationship between wind speed and the theoretical output of a single wind turbine within the range from cut-in wind speed, rated wind speed to cut-out wind speed, thus converting the wind speed value into the theoretical output value of a single wind turbine. The theoretical output values ​​of all wind turbines within a single wind farm are summed to obtain the total theoretical output value of that wind farm. The total theoretical output values ​​of all wind farms within the target area are further summarized to obtain the total theoretical wind power output value of the target area for that hour. Data characterizing the wind field turbulence intensity is extracted from the generated basic meteorological element field. Turbulence intensity reflects the fluctuation of wind speed. A correction factor is calculated based on the turbulence intensity data; this correction factor is inversely proportional to the turbulence intensity. This correction factor is used to reduce the total theoretical wind power output to account for the adverse effects of turbulence on the overall wind farm output. The hourly total theoretical wind power output, after reduction, is concatenated in chronological order to form a wind power output response curve. Hydrological forecast data for the watershed where the hydropower station is located within the target area is obtained. This data includes the predicted inflow and water level of the upstream reservoir. The hourly relative humidity distribution raster is fused with the watershed hydrological forecast data to identify areas where relative humidity remains consistently high and may trigger heavy rainfall, along with their duration. Based on the spatial coverage and duration of the identified high relative humidity areas, the predicted inflow of the upstream reservoir is corrected, adding the additional inflow that may be due to predicted rainfall, generating a corrected inflow prediction. The corrected inflow prediction and the predicted water level are then input into the hydropower station's output calculation formula, which typically involves parameters such as flow rate, head, and efficiency, to calculate the corrected output of the hydropower station considering the impact of rainfall. The calculated corrected output value is compared with a baseline output value derived solely from the reservoir scheduling plan without considering the impact of precipitation. The resulting ratio is the hydropower output correction coefficient.

[0027] In practical implementation, the basic meteorological element field includes hourly temporal resolution grids for temperature distribution, relative humidity distribution, and surface wind speed distribution in East China as of July 15, 2026. The spatial resolution of the temperature distribution grid is 5 kilometers, with a total of 10,000 grid points. From the hourly temporal resolution temperature distribution grid, based on land use type data from the geographic information system, the temperature values ​​of all grid points labeled "residential" are extracted and averaged to obtain the hourly average temperature for residential areas. Similarly, the temperature values ​​of all grid points labeled "commercial" are extracted and averaged to obtain the hourly average temperature for commercial areas. The preset temperature comfort range is 26 to 28 degrees Celsius, with an upper comfort limit of 28 degrees Celsius and a lower comfort limit of 26 degrees Celsius. The calculations are based on the average temperature of the residential area and the preset temperature comfort range. When the average temperature of the residential area is higher than 28 degrees Celsius, the difference between it and 28 degrees Celsius is calculated; when the average temperature of the residential area is lower than 26 degrees Celsius, the difference between it and 26 degrees Celsius is calculated. The average temperature of the residential area at 14:00 on the predicted day is 32 degrees Celsius, which is higher than the upper limit of comfort by 28 degrees Celsius, with a difference of 4 degrees Celsius. The cooling demand index of the residential area is the cumulative value of the differences in temperature above the upper limit of comfort throughout the 24 hours, summing up all differences above 28 degrees Celsius. The heating demand index of the residential area is the cumulative value of the differences in temperature below the lower limit of comfort throughout the 24 hours, summing up all differences below 26 degrees Celsius. The preset commercial temperature sensitivity coefficient is 150 megawatts per degree Celsius. The air conditioning load adjustment of the commercial area is equal to the deviation of the commercial area temperature from the median of the comfort range multiplied by the commercial temperature sensitivity coefficient. The median of the comfort range is 27 degrees Celsius. The predicted average temperature in the commercial area at 2 PM is 30 degrees Celsius, with a deviation of 3 degrees Celsius. The air conditioning load adjustment for the commercial area is 3 degrees Celsius multiplied by 150 megawatts per degree Celsius, resulting in 450 megawatts.

[0028] In some embodiments, turbulence intensity data related to wind power output reduction is extracted from the basic meteorological data field. Turbulence intensity is a physical quantity that measures the severity of wind speed fluctuations over time; its standard definition is the ratio of the standard deviation of the wind speed time series at a specific location to the average wind speed over that time period. The regional meteorological model provides hourly forecasts of wind speed time series for each wind farm location, or directly outputs turbulent kinetic energy fields. The turbulence intensity data is calculated by processing the forecast wind speed time series for each location, or converted from the turbulent kinetic energy field output by the model. The correction factor used for the reduction calculation is inversely proportional to the turbulence intensity value obtained in this manner. The total theoretical wind power output is reduced by multiplying the total theoretical wind power output by a correction factor inversely proportional to the turbulence intensity. The formula for calculating the correction factor is: ; Where: η represents the correction factor, α represents the attenuation coefficient with a value of 0.5, and I represents the turbulence intensity. When the turbulence intensity I is 0.1, the correction factor η is 1 / (1+0.5*0.1), approximately equal to 0.952. The total theoretical wind power output is 800 MW, which, after reduction, is 800 MW multiplied by 0.952, resulting in 761.6 MW. It can be understood that connecting the reduced hourly total theoretical wind power output values ​​in chronological order forms a wind power output response curve. The wind power output response curve is a time series containing 24 data points, each corresponding to a one-hour reduced predicted wind power output value.

[0029] In practice, the watershed hydrological forecast data for hydropower stations within the target area is acquired. This data includes the predicted inflow and water level of the upstream reservoir for the next 24 hours. The predicted inflow is recorded hourly, in cubic meters per second. The predicted water level is also recorded hourly, in meters. The hourly relative humidity distribution grid is fused with the watershed hydrological forecast data to identify areas of high relative humidity that trigger heavy rainfall and their duration. A relative humidity threshold of 85% is set. Areas within the upstream catchment area of ​​the reservoir where the relative humidity value exceeds 85% for three consecutive hours are identified, covering an area of ​​200 square kilometers. Based on the coverage and duration of the high relative humidity area, the predicted inflow of the upstream reservoir is corrected. The correction formula adds an increment proportional to the area and duration of the high humidity region to the original predicted inflow. A revised inflow forecast is generated, increasing the original forecast of 1000 cubic meters per second by 50 cubic meters per second, resulting in 1050 cubic meters per second. Substituting the revised inflow forecast and the reservoir water level forecast into the hydropower station's output calculation formula, the hydropower station's output calculation formula is as follows: ; Where: symbol P represents the power output of the hydropower station; symbol The efficiency of the hydroelectric generator unit is represented by a value of 0.9; the symbol H represents the head corresponding to the reservoir water level; the symbol Q represents the inflow rate in cubic meters per second; the constant 9.81 is the specific weight of water, approximately 9.81 kN / m³. The corrected output value considering the impact of precipitation is calculated. This corrected output value is then compared with the baseline output value excluding the impact of precipitation. The baseline output value excluding the impact of precipitation is calculated based on the original predicted inflow rate of 1000 cubic meters per second. The corrected output value is calculated based on the corrected predicted inflow rate of 1050 cubic meters per second. The result of the ratio calculation is the corrected output value divided by the baseline output value, yielding the hydroelectric output correction coefficient.

[0030] In one embodiment of the present invention, see [reference] Figure 3 The air conditioning load response curve, wind power output response curve, and hydropower output correction coefficient are mapped to the same time coordinate system with the hour as the smallest unit, ensuring strict alignment of the three curves on the time axis. The hydropower output correction coefficient is applied to a preset hydropower baseline output curve, which represents the theoretical output curve of the hydropower station under conditions of no precipitation, based on the reservoir scheduling plan. This is achieved by multiplying the hourly baseline hydropower output value by the corresponding hourly hydropower output correction coefficient, thus generating a corrected hydropower output curve that considers the effects of humidity and precipitation. The corrected hydropower output curve is then added to the wind power output response curve hourly. Based on this, baseline output prediction values ​​for other power sources such as thermal power and nuclear power are added. These baseline output prediction values ​​can be determined based on factors such as unit maintenance plans and fixed output agreements. The output values ​​of all power sources are summarized hourly to generate an energy supply-side output prediction dataset covering all time points within the prediction period. The air conditioning load response curve is added hourly to the predicted values ​​of other rigid loads besides air conditioning load. These other rigid load predictions can be based on non-meteorologically sensitive factors such as historical load data, weekday patterns, and holiday information. The summaries generate an energy consumption-side load prediction dataset covering all time points within the prediction period. Both the generated energy supply-side output prediction dataset and the energy consumption-side load prediction dataset are then smoothed and filtered. Filters are applied to remove abnormal fluctuations introduced by random noise during data acquisition that do not conform to the physical characteristics of the power system, making the prediction curves smoother and more reasonable.

[0031] In practice, the generated energy supply-side output forecast dataset and energy consumption-side load forecast dataset are both subjected to smoothing filtering. The smoothing filtering employs a simple moving average method with a length of 3 to remove abnormal fluctuations introduced by data acquisition noise. The formula for calculating the moving average is: ; Where: symbol This represents the value after smoothing and filtering at time index t, with the sign... Represents the original value at time index t-1, symbol Represents the original value at time index t, symbol This represents the original value at time index t+1. For example, for the energy supply-side output forecast dataset, the original value at time t=2 is 2616.2 MW, and its smoothed value is... The calculated value is (2500 + 2616.2 + 2736) / 3, yielding 2617.4 MW. The original value of the energy consumption-side load forecast dataset at time t=2 is 1895 MW, and its smoothed value is... The result is (1950 + 1895 + 1840) / 3, which yields 1895 MW. At the start and end points of the time series (t=1 and t=24), a boundary replication method is used for smoothing, employing the most recent effective values. After smoothing filtering, the energy supply-side output forecast dataset becomes [2500, 2617.4, 2730.7, ..., 2354.4] MW. The energy consumption-side load forecast dataset becomes [1950, 1895, 1861.7, ..., 2000] MW.

[0032] In one embodiment of the invention, the data points for each hour in the energy supply-side output forecast dataset are subtracted from the corresponding data points for the same hour in the energy consumption-side load forecast dataset. The result is the net power surplus or net power deficit value for that hour. A supply-demand balance threshold is set, representing a small range of power fluctuations acceptable to the power system that does not require trading adjustments. The calculated hourly net power surplus or deficit value is judged; if its absolute value is less than the set threshold, the value for that hour is reset to zero, indicating that the power system is in a supply-demand balance state for that hour. The sequence of net power surplus or deficit values ​​after zeroing is accumulated and integrated. Starting from the beginning of the forecast period, the processed values ​​for each hour are summed to generate a cumulative supply-demand deficit curve. This cumulative supply-demand deficit curve reflects the cumulative degree of power supply-demand imbalance over a period of time. The hourly net power surplus or deficit value is used as instantaneous information and co-encoded with the cumulative information represented by the cumulative supply and demand gap curve. For example, a data sequence containing three fields is formed, namely "time stamp, instantaneous gap value, and cumulative gap value". Finally, a time series of power trading supply and demand gap is generated that includes both instantaneous power trading demand and reflects the cumulative trend of supply and demand contradiction.

[0033] In practice, the energy supply-side output forecast dataset is a time series containing 24-hour forecasts from 00:00 to 23:00 on August 10, 2026, while the energy consumption-side load forecast dataset is a time series with the same date and duration. The net power surplus or net power deficit for each hour is calculated by subtracting the corresponding hourly data point from the energy consumption-side load forecast dataset. At time index t=1, the energy supply-side output forecast is 2550 MW, and the energy consumption-side load forecast is 2450 MW; the subtraction result is +100 MW, indicating a net power surplus of 100 MW. At time index t=14, the energy supply-side output forecast is 3200 MW, and the energy consumption-side load forecast is 3350 MW; the subtraction result is -150 MW, indicating a net power deficit of 150 MW.

[0034] A supply-demand balance threshold of 50 MW is set. When the absolute value of the net power surplus or deficit is less than the threshold of 50 MW, the corresponding hourly value is reset to zero, indicating a supply-demand balance. At time index t=7, the calculated net power surplus is +30 MW, and its absolute value of 30 MW is less than the threshold of 50 MW, so this value is reset to zero. At time index t=20, the calculated net power deficit is -40 MW, and its absolute value of 40 MW is also less than the threshold of 50 MW, so it is also reset to zero. For values ​​with an absolute value greater than or equal to 50 MW, the original calculation result is retained. The sequence of net power surplus or deficit values ​​after zeroing reflects the periods of significant imbalance requiring market transaction adjustment. See Table 1.

[0035] Table 1: Calculation Table of Electricity Supply and Demand Gap Time index (t) Time point Energy supply-side output forecast (megawatts) Energy consumption-side load forecast (megawatts) Original net electricity surplus / deficit value (megawatts) Determine the processed value (megawatts). Cumulative supply-demand gap (megawatts) 1 00:00 2550 2450 +100 +100 +100 7 06:00 2800 2770 +30 0 +400 14 13:00 3200 3350 -150 -150 +100 20 19:00 3100 3060 +40 0 -10 24 23:00 2600 2620 -20 0 -30 The net electricity surplus or deficit value after zeroing is accumulated and integrated to generate a cumulative supply-demand gap curve. The cumulative supply-demand gap curve starts from the prediction start time and accumulates the processed values ​​at each time point. The cumulative calculation formula is: ; Where: symbol This represents the cumulative supply-demand gap value up to time index t, with the symbol... This represents the net power surplus or deficit value after judgment and zeroing at time index k. The processed value at time index t=1. +100 MW, cumulative value It is +100 megawatts. At time index t=7, due to 0 megawatts, cumulative value for Add 0, assuming If it is +400 megawatts, then It remains at +400 MW. The cumulative supply-demand gap reflects the cumulative deviation between electricity supply and demand over a period of time; a positive value indicates a cumulative electricity surplus, while a negative value indicates a cumulative electricity shortage.

[0036] In some embodiments, the net electricity surplus or deficit value is co-encoded with the cumulative supply-demand deficit curve to generate a power trading supply-demand deficit time series containing instantaneous deficit information and cumulative deficit information. The power trading supply-demand deficit time series is a data structure containing multiple records, each corresponding to a point in time. Each record contains three core fields: timestamp, instantaneous deficit value, and cumulative deficit value. The timestamp field is accurate to the hour. The instantaneous deficit value field stores the net electricity surplus or deficit value after threshold processing and zeroing, as shown in the "Value after Threshold Processing" column of Table 1. The cumulative deficit value field stores the cumulative supply-demand deficit value up to that point, as shown in the "Cumulative Supply-Demand Deficit Value" column of Table 1.

[0037] See Figure 4 This is a time-series comparison chart of electricity supply and demand, showing the 24-hour trends of energy supply-side output and energy consumption-side load from 00:00 to 23:00 on August 10, 2026. It serves as a visualization of fundamental data for electricity trading supply and demand forecasting. Both curves exhibit a typical daily load / output pattern of "rising from early morning to midday and falling back from afternoon to evening." The peaks on both the supply and consumption sides occur around 13:00, reaching approximately 3200 MW and 3350 MW respectively. At 00:00, the supply side has 2550 MW and the consumption side has 2450 MW, resulting in a surplus of 100 MW. At 06:00, the supply side has 2800 MW and the consumption side has 2770 MW, resulting in a surplus of 30 MW. At 13:00, the supply side has 3200 MW and the consumption side has 3350 MW, resulting in a deficit of 150 MW. At 19:00, the supply side had 3100 MW and the consumption side had 3060 MW, resulting in a surplus of 40 MW. At 23:00, the supply side had 2600 MW and the consumption side had 2620 MW, resulting in a deficit of 20 MW. For most of the time, the supply side was slightly higher than the consumption side, showing a small surplus. A significant deficit appeared around 13:00, which was the period of tightest power supply and demand for the day.

[0038] In one embodiment of the present invention, the system accesses real-time short-term and nowcast data from a regional meteorological big data model. This data includes hourly updated meteorological element forecasts for the next 24 hours starting at 12:00 on August 10, 2026. The meteorological element forecasts include the latest grid predictions for temperature, surface wind speed, and relative humidity, with a spatial resolution of 5 kilometers. At 12:00, the regional meteorological big data model outputs an updated forecast, showing that the average temperature forecast for the target area from 13:00 to 14:00 has been adjusted from 32 degrees Celsius to 34 degrees Celsius, and the average wind speed forecast has been adjusted from 6.5 meters per second to 5.0 meters per second. This short-term and nowcast data is input into the rolling forecast interface of the regional meteorological big data model, completely replacing the medium- and long-term meteorological forecast data originally covering August 10 and 11 used in the first round of forecasts at 00:00 on August 10 with the 24-hour forecast data updated at 12:00 on August 10, 2026. The incremental learning mechanism in the regional meteorological big data model is triggered, using the latest real-time power load data and real-time renewable energy generation data to fine-tune the model weights. The latest real-time power load data is the actual load value of 3050 MW at 11:00 on August 10, 2026, and the latest real-time renewable energy generation data is the actual wind power generation value of 580 MW at 11:00 on August 10, 2026. The purpose of the incremental learning mechanism is to enable the regional meteorological big data model to quickly adapt to the latest real-time meteorological trends and system operating status.

[0039] The fine-tuned regional meteorological model is invoked to re-execute the forecasting process from generating the basic meteorological element field to generating the time series of electricity trading supply and demand gap. Based on the new meteorological inputs and the latest system state, the fine-tuned model regenerates the air conditioning load response curve, wind power output response curve, and hydropower output correction coefficient for the next 24 hours, thereby obtaining updated energy supply-side output forecast datasets and energy consumption-side load forecast datasets. Finally, a new time series of electricity trading supply and demand gap is calculated. The regenerated time series of electricity trading supply and demand gap is compared with the previous forecast results, and the revised incremental data of supply and demand forecast is output. The previous forecast gave an electricity trading supply and demand gap of -150 MW for the 13:00-14:00 period. The regenerated forecast gives an electricity trading supply and demand gap of -200 MW for the 13:00-14:00 period. The revised incremental data is the difference between the two: -200 - (-150) = -50 MW. This -50 MW revised increment indicates that, based on the latest meteorological information, the power shortage expected between 13:00 and 14:00 is 50 MW more than previously predicted.

[0040] In some embodiments, the process of fine-tuning the model weights using the latest real-time power load data and real-time renewable energy generation data includes several steps: calculating the deviation rate between the latest real-time power load data and the predicted load value at the previous time step, and the deviation rate between the latest real-time renewable energy generation data and the predicted generation value at the previous time step. The predicted load value at the previous time step was 2980 MW, the real-time load value was 3050 MW, and the load deviation rate was (3050-2980) / 2980≈2.35%. The predicted wind power generation value at the previous time step was 620 MW, the real-time wind power generation value was 580 MW, and the generation deviation rate was (580-620) / 620≈-6.45%. The load deviation rate is compared with a preset alarm threshold of 2.0%. The generation deviation rate is compared with a preset alarm threshold of 5.0%. Abnormal data points with deviation rates exceeding the alarm thresholds are filtered out. The load deviation rate of 2.35% is greater than the alarm threshold of 2.0%, therefore the load data point at 11:00 is filtered as an abnormal data point. The absolute value of the power generation deviation rate of -6.45% is greater than the alarm threshold of 5.0%, therefore the wind power generation data point at 11:00 is also filtered as an abnormal data point.

[0041] The spatiotemporal features corresponding to the abnormal data points are extracted and input as negative samples into the loss function of the regional meteorological big data model. Before being input into the loss function, these real-time spatiotemporal features need to be reconstructed into a feature vector format consistent with the input dimension of the model's multi-source data fusion layer. Specifically, real-time meteorological data and load and power generation data are calculated in real time according to the same feature extraction rules as in the model pre-training stage to generate real-time load-side and power-side meteorological correlation feature vectors, which are then input into the model for weight adjustment. For the load abnormal data point at 11:00, the extracted spatiotemporal features include: time feature "11:00", location feature "the whole province", and corresponding real-time meteorological features "temperature 31 degrees Celsius, wind speed 6.2 meters per second". The same spatiotemporal features are extracted for the wind power abnormal data point at 11:00. These feature vectors are combined into negative samples and input into the loss function, which calculates the error between the current model's prediction based on these features and the actual observation value. The weights of the fully connected layers related to the mapping of meteorological elements in the regional meteorological big data model are adjusted through the backpropagation algorithm to reduce the model's confidence in the meteorological features that caused the abnormal data points. The backpropagation algorithm updates the weight parameters of the fully connected layers based on the gradient calculated from the loss function. For example, for load forecasting, the model might reduce the estimated weight of the impact of the feature "wind speed of 6.2 m / s" on load during the midday period. For wind power forecasting, the model might adjust the weight of the mapping relationship between "temperature of 31 degrees Celsius" and wind speed on power output. The process of calculating the deviation rate, filtering outlier data points, inputting negative samples, and adjusting the weights of the fully connected layers is repeated until the deviation rate converges to within the alarm threshold, completing the model fine-tuning. At the next rolling time of 12:00, real-time data from 11:00 to 12:00 is acquired again, and a new deviation rate is calculated. After the previous round of weight adjustments, the new load forecast value might be 3040 MW, the real-time value is 3055 MW, and the deviation rate is approximately 0.49%, lower than the 2.0% alarm threshold. The new wind power forecast value might be 585 MW, the real-time value is 590 MW, and the deviation rate is approximately 0.85%, lower than the 5.0% alarm threshold. When the deviation rate is lower than the alarm threshold for multiple consecutive time periods, it is determined that the deviation rate has converged, the incremental learning mechanism pauses weight adjustment, and the model fine-tuning is completed.

[0042] See Figure 5This is a comparison chart of the rolling correction effect of load forecasting, showing the changing trends of the previous round of forecasted load, actual load, and the fine-tuned forecasted load over 24 hours. It visually demonstrates the improvement in forecast accuracy achieved through model fine-tuning. The actual load curve remains at the top, indicating that the overall forecast value is slightly lower than the actual value. The fine-tuned forecasted load shows a significantly higher degree of alignment with the actual load than the previous forecast, especially during the peak period from 12:00 to 22:00, where the deviation is noticeably reduced. During the off-peak period from 23:00 to 06:00, the fine-tuned forecast curve also more closely reflects the fluctuations in actual load. From 17:00 to 18:00: the actual load peak was approximately 3150 MW, while the fine-tuned forecast was approximately 3130 MW, a deviation of only about 0.6%; whereas the previous forecast was approximately 3040 MW, a deviation of about 3.5%. 06:00: The actual load is about 2550 MW, and the slightly adjusted forecast is about 2530 MW, with a deviation of about 0.8%; while the previous forecast was about 2450 MW, with a deviation of about 3.9%.

[0043] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for predicting power trading supply and demand based on a regional meteorological large-scale model, characterized in that, The method includes: Historical power load data, historical renewable energy generation data, and historical meteorological observation data of the target area are acquired and input into a pre-trained regional meteorological big model. The multi-source data fusion layer in the regional meteorological big model performs feature cross-calculation on the historical power load data, historical renewable energy generation data, and historical meteorological observation data to generate a basic meteorological element field of the target area. The basic meteorological element field includes a temperature distribution grid, a relative humidity distribution grid, and a surface wind speed distribution grid with a time resolution of hour. Based on the aforementioned basic meteorological element field, the energy response mapping module in the regional meteorological big model is invoked to map the hourly temperature distribution grid as an air conditioning load response curve, the hourly surface wind speed distribution grid as a wind power output response curve, and the hourly relative humidity distribution grid as a hydropower output correction coefficient. The air conditioning load response curve, wind power output response curve and hydropower output correction coefficient are time-series aligned to generate the energy supply-side output prediction dataset and the energy consumption-side load prediction dataset for the target area. Based on the difference between the energy supply-side output forecast dataset and the energy consumption-side load forecast dataset, a time series of the power trading supply and demand gap in the target region is generated.

2. The power trading supply and demand forecasting method based on a regional meteorological large model according to claim 1, characterized in that, The data is input into a pre-trained regional meteorological model. The historical power load data, historical renewable energy generation data, and historical meteorological observation data are then cross-calculated using the multi-source data fusion layer within the regional meteorological model to generate a basic meteorological element field for the target region, including: The historical power load data includes power consumption curves for various industries; the historical new energy power generation data includes irradiance conversion curves for photovoltaic power plants and wind speed power curves for wind turbines; and the historical meteorological observation data includes spatiotemporal distribution records of temperature, humidity, air pressure, and cloud cover. The historical meteorological observation data were subjected to quality control processing to remove sensor drift data and missing period data, and to complete the time series of missing period data. The historical power load data is divided into industrial load subsequence, commercial load subsequence and residential load subsequence according to industry category, and the historical new energy power generation data is divided into photovoltaic power output subsequence and wind power output subsequence. The industrial load subsequence, commercial load subsequence, and residential load subsequence are correlated with the historical meteorological observation data with a time resolution of hours to extract the meteorological sensitive feature vector on the load side. The photovoltaic power output subsequence and the wind power output subsequence are coupled with the historical meteorological observation data with a time resolution of hours for physical mechanism analysis to extract the meteorological driving feature vector on the power supply side. The meteorological sensitive feature vector from the load side and the meteorological driving feature vector from the power supply side are input into the multi-source data fusion layer of the regional meteorological big model. The data is then processed by a multilayer perceptron network to perform nonlinear transformation, and the output is the basic meteorological element field containing the spatial distribution characteristics of air temperature, relative humidity, and surface wind speed.

3. The power trading supply and demand forecasting method based on a regional meteorological large model according to claim 2, characterized in that, Mapping the temperature distribution grid with an hourly time resolution to an air conditioning load response curve includes: The average temperature of the residential area and the average temperature of the commercial area are extracted from the temperature distribution grid with a time resolution of hours. Based on the average temperature of the residential area and the preset temperature comfort range, the cooling demand index and heating demand index of the residential area are calculated. The cooling demand index is the cumulative value of the difference between the residential area temperature and the upper limit of comfort, and the heating demand index is the cumulative value of the difference between the residential area temperature and the lower limit of comfort. Based on the average temperature of the commercial area and the preset commercial temperature sensitivity coefficient, the air conditioning load adjustment of the commercial area is calculated. The air conditioning load adjustment of the commercial area is equal to the deviation between the temperature of the commercial area and the median of the comfort range multiplied by the commercial temperature sensitivity coefficient. The total air conditioning load forecast value of the target area is generated by weighted summing of the cooling demand index and heating demand index of the residential area and the air conditioning load adjustment of the commercial area. The total air conditioning load forecast values ​​are connected in hourly order to form the air conditioning load response curve.

4. The power trading supply and demand forecasting method based on a regional meteorological large model according to claim 3, characterized in that, Mapping the surface wind speed distribution grid with a time resolution of hours to a wind power output response curve includes: Spatial downscaling is performed on the surface wind speed distribution grid with a time resolution of hour to convert the low-resolution grid data into a high-resolution wind speed sequence corresponding to the wind farm turbine location coordinates. For each wind speed value in the high-resolution wind speed sequence, a preset wind turbine power curve lookup table is called to convert the wind speed value into the theoretical output value of a single wind turbine. The wind turbine power curve lookup table contains the corresponding power relationship between the cut-in wind speed, the rated wind speed, and the cut-out wind speed. The theoretical output values ​​of all wind turbines in a single wind farm are summed to obtain the total theoretical output value of the single wind farm. The total theoretical output values ​​of all wind farms in the target area are then summarized to obtain the total theoretical wind power output value of the target area. Based on the turbulence intensity data of the surface wind speed distribution grid extracted from the basic meteorological element field, the total theoretical wind power output is reduced by multiplying the total theoretical wind power output by a correction factor that is inversely proportional to the turbulence intensity. The reduced total theoretical wind power output is connected in hourly order to form the wind power output response curve.

5. The power trading supply and demand forecasting method based on a regional meteorological large model according to claim 4, characterized in that, Mapping the relative humidity distribution grid with a time resolution of hours to a hydropower output correction factor includes: Obtain the watershed hydrological forecast data of hydropower stations within the target area, wherein the watershed hydrological forecast data includes the predicted inflow and water level of upstream reservoirs; By fusing and analyzing the relative humidity distribution grid with a time resolution of hours with the watershed hydrological forecast data, the high relative humidity areas that trigger heavy precipitation and their duration can be identified. Based on the coverage and duration of the high relative humidity area, the predicted inflow of the upstream reservoir is corrected to generate a corrected predicted inflow. Substituting the corrected inflow forecast and reservoir water level forecast into the power output calculation formula of the hydropower station, the corrected power output value considering the impact of precipitation is calculated. The ratio of the corrected output value to the baseline output value (which does not consider the impact of precipitation) is the result of the hydropower output correction coefficient.

6. The power trading supply and demand forecasting method based on a regional meteorological large model according to claim 5, characterized in that, The air conditioning load response curve, wind power output response curve, and hydropower output correction coefficient are time-series aligned to generate an energy supply-side output prediction dataset and an energy consumption-side load prediction dataset for the target area, including: The air conditioning load response curve, wind power output response curve, and hydropower output correction coefficient are mapped to the same time coordinate system, with the smallest unit of the time coordinate system being the hour. The hydropower output correction coefficient is applied to the preset hydropower reference output curve to generate the corrected hydropower output curve. The hydropower reference output curve is the output curve of the hydropower station under the condition of no precipitation according to the reservoir scheduling plan. The modified hydropower output curve is added to the wind power output response curve, and the baseline output prediction values ​​of thermal power and other power sources are added to generate the energy supply-side output prediction dataset. The air conditioning load response curve is added to the predicted values ​​of other rigid loads besides the air conditioning load, and the data is summarized to generate the energy consumption side load prediction dataset. The generated energy supply-side output forecast dataset and energy consumption-side load forecast dataset are respectively subjected to smoothing filtering to remove abnormal fluctuation points introduced by data acquisition noise.

7. The power trading supply and demand forecasting method based on a regional meteorological large model according to claim 6, characterized in that, Based on the difference between the energy supply-side power output forecast dataset and the energy consumption-side load forecast dataset, a time series of the power trading supply-demand gap in the target region is generated, including: Subtract the corresponding hourly data point from the energy supply-side output forecast dataset to obtain the net power surplus or deficit value for each hour. A threshold for determining supply and demand balance is set. When the absolute value of the net power surplus or deficit is less than the threshold, the value of the corresponding hour is set to zero, which is regarded as a state of supply and demand balance. The net power surplus or deficit value after zeroing is accumulated and integrated to generate a cumulative supply and demand gap curve, which reflects the cumulative deviation of power supply and demand over a period of time. The net power surplus or deficit value is encoded together with the cumulative supply and demand deficit curve to generate the power trading supply and demand deficit time series containing instantaneous deficit information and cumulative deficit information.

8. The power trading supply and demand forecasting method based on a regional meteorological large model according to claim 7, characterized in that, It also includes a rolling revision step for supply and demand forecasts based on real-time weather updates: The system accesses the short-term and nowcast data output by the regional meteorological big model, which includes hourly updated meteorological element forecasts for the next 24 hours. The short-term and nowcast data is input into the rolling forecast interface of the regional meteorological big model to replace the original medium- and long-term meteorological forecast data. The incremental learning mechanism in the regional meteorological big data model is triggered, and the model weights are fine-tuned using the latest real-time power load data and real-time new energy power generation data to adapt to the latest meteorological change trends. The fine-tuned regional meteorological model is invoked to re-execute the forecasting process from generating the basic meteorological element field to generating the time series of electricity trading supply and demand gap; The regenerated time series of power trading supply and demand gap is compared with the previous forecast results using differential analysis, and the revised incremental data of supply and demand forecast is output.

9. A power trading supply and demand forecasting method based on a regional meteorological large model according to claim 8, characterized in that, The method of fine-tuning the model weights using the latest real-time power load data and real-time renewable energy generation data to adapt to the latest meteorological trends includes: Calculate the deviation rate between the latest real-time power load data and the predicted load value at the previous moment, and the deviation rate between the latest real-time renewable energy generation data and the predicted power generation value at the previous moment; The deviation rate is compared with a preset alarm threshold to filter out abnormal data points whose deviation rate exceeds the alarm threshold. Extract the spatiotemporal features corresponding to the abnormal data points and input them as negative samples into the loss function of the regional meteorological big model; By adjusting the weights of the fully connected layers related to the mapping of meteorological elements in the regional meteorological big model through the backpropagation algorithm, the confidence of the model in the meteorological characteristics that lead to the abnormal data points is reduced. The process of calculating the repetition rate, filtering outlier data points, inputting negative samples, and adjusting the weights of the fully connected layers continues until the repetition rate converges to within the alarm threshold, thus completing the fine-tuning of the model.

10. A power trading supply and demand forecasting system based on a regional meteorological large model, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the power trading supply and demand forecasting method based on a regional meteorological big data model as described in any one of claims 1 to 9.

Images