Method and system for intelligent prediction of power load based on meteorological evaluation

Abstract

The application discloses a power load intelligent prediction method and system based on meteorological evaluation, and relates to the technical field of power load prediction. The method comprises the following steps: collecting multi-source prediction data of a target area in a preset time window, performing standardization processing and space-time alignment, and generating a multi-dimensional input tensor; performing load prediction complexity evaluation according to the multi-source prediction data, determining a load prediction complexity index, calling an adaptive load predictor, and predicting a power load prediction result according to the multi-dimensional input tensor; determining whether the target area in the preset time window is non-extreme weather or extreme weather based on the multi-source prediction data, and if it is extreme weather, calling an extreme weather correction model to correct the power load prediction result and output a corrected load prediction result. Through load prediction complexity evaluation, the application adaptively calls a prediction model, and performs targeted correction under extreme weather, so that the adaptive balance of prediction accuracy and calculation efficiency is realized.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and specifically to a method and system for intelligent power load forecasting based on meteorological assessment. Background Technology

[0002] Electricity load forecasting is the fundamental support for power grid dispatching and supply guarantee decisions. With the increasing proportion of renewable energy integration and the frequent occurrence of extreme weather, load forecasting faces greater uncertainty. On the one hand, the volatility of renewable energy output such as wind power and photovoltaics increases the uncertainty of net load forecasting; on the other hand, extreme weather events such as extreme high temperatures and cold waves can cause drastic load fluctuations, and traditional forecasting methods are unable to accurately capture the load change patterns under such extreme scenarios.

[0003] Existing technologies mainly employ forecasting methods driven by general weather forecasts and shallow statistical modeling methods based on historical time series. The former uses weather forecasts as exogenous variables combined with machine learning for prediction, while the latter utilizes models such as ARIMA or decision trees to model historical load sequences. However, existing technologies are insufficient in characterizing small-scale urban meteorological differences and sudden weather events, struggle to integrate the complex mapping relationship between multi-source meteorological information and loads, and cannot adaptively adjust forecasting strategies according to the complexity of meteorological changes. Furthermore, they lack effective correction mechanisms in extreme weather scenarios, leading to a significant increase in forecasting errors.

[0004] Therefore, there is an urgent need for a load forecasting method that can adaptively invoke forecasting models based on meteorological complexity and make effective corrections under extreme weather conditions. Summary of the Invention

[0005] This invention addresses the technical problems in existing technologies, such as the difficulty in integrating the complex mapping relationship between multi-source meteorological information and load, the inability to adaptively adjust prediction strategies according to the complexity of meteorological changes, and the lack of effective correction mechanisms in extreme weather scenarios. It provides a method and system for intelligent power load prediction based on meteorological assessment.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for intelligent forecasting of power load based on meteorological assessment, comprising: Collect multi-source prediction data of the target area within a preset time window, and perform standardization and spatiotemporal alignment on the multi-source prediction data to generate a multi-dimensional input tensor; Based on the multi-source forecast data, the load forecast complexity index is determined by assessing the load forecast complexity. The load forecast complexity index is then used to call the adaptive load forecaster. Based on the multi-dimensional input tensor, the power load forecast result within the preset time window is obtained. Based on the multi-source prediction data, it is determined whether the target area will experience non-extreme or extreme weather within a preset time window. If it is non-extreme weather, the power load prediction result is output. If it is extreme weather, the extreme weather correction model is called to correct the power load prediction result, and the corrected load prediction result is output.

[0007] Secondly, the present invention provides an intelligent power load forecasting system based on meteorological assessment, comprising: The data acquisition and processing module is used to acquire multi-source prediction data of the target area within a preset time window, perform standardization and spatiotemporal alignment on the multi-source prediction data, and generate a multi-dimensional input tensor. The load forecasting module is used to assess the load forecasting complexity based on the multi-source forecasting data, determine the load forecasting complexity index, call the adaptive load forecaster based on the load forecasting complexity index, and predict the power load forecasting result within a preset time window based on the multi-dimensional input tensor. The weather determination and correction module is used to determine whether the target area will experience non-extreme or extreme weather within a preset time window based on the multi-source prediction data. If it is non-extreme weather, the module outputs the power load prediction result. If it is extreme weather, the module calls the extreme weather correction model to correct the power load prediction result and outputs the corrected load prediction result.

[0008] The beneficial effects of this invention are: Compared to existing technologies, this invention first constructs a unified multidimensional input tensor by standardizing and spatiotemporally aligning multi-source forecast data, providing a high-quality, semantically consistent data foundation for load forecasting models. Secondly, it determines a load forecasting complexity index through load forecasting complexity assessment and dynamically calls appropriate load forecasters based on this index, matching the complexity of the forecasting model to the severity of weather changes. More forecasting units are called during periods of severe weather fluctuations to improve accuracy, while fewer units are called during periods of stable weather to reduce computational overhead. Thirdly, an extreme weather determination mechanism distinguishes between normal and extreme scenarios. In normal scenarios, forecast results are directly output to ensure efficiency, while in extreme scenarios, an extreme weather correction model is invoked to specifically correct the forecast results, compensating for the performance degradation of normal forecasting models under extreme weather conditions. This invention solves the problems of fixed and singular forecasting strategies and large forecasting errors under extreme weather conditions in existing technologies, achieving an adaptive balance between load forecasting accuracy and computational efficiency. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the intelligent power load forecasting method based on meteorological assessment provided by the present invention. Figure 2 This is a schematic diagram of the intelligent power load forecasting system based on meteorological assessment provided by the present invention.

[0010] In the attached diagram, the components represented by each number are as follows: Data acquisition and processing module 11, load forecasting module 12, weather determination and correction module 13. Detailed Implementation

[0011] Example 1, as Figure 1 As shown, this embodiment of the invention provides a smart power load forecasting method based on meteorological assessment, including: S10: Collect multi-source prediction data of the target area within a preset time window, perform standardization processing and spatiotemporal alignment on the multi-source prediction data, and generate a multi-dimensional input tensor; First, multi-source forecast data is collected for the target area within a preset time window. The target area refers to the power grid jurisdiction that requires load forecasting, such as the administrative region covered by a prefecture-level power grid. Within this target area, load distribution is affected by various factors such as meteorological conditions and renewable energy access, requiring forecast information to be obtained from multiple data sources. The preset time window refers to the future time period for which load forecasting is required, such as the next 24 hours or the next 48 hours. The length of this window can be set according to the needs of power grid dispatching operations. For example, day-ahead dispatching requires load forecasting for the next 24 hours, while intraday rolling dispatching requires load forecasting for the next 4 to 8 hours.

[0012] The multi-source forecast data includes meteorological forecast data, historical load data, and expected energy type characteristic data. The meteorological forecast data includes one or more of temperature data, humidity data, wind speed data, air pressure data, and irradiance data. The historical load data includes the electricity load sequence of the target area within a historical time range. The expected energy type characteristic data includes one or more of wind power access ratio data, photovoltaic access ratio data, and hydropower access ratio data.

[0013] Specifically, multi-source forecast data includes meteorological forecast data, historical load data, and characteristic data of expected energy types. Meteorological forecast data comes from numerical weather prediction systems and is used to describe changes in atmospheric conditions within a preset future time window. This includes one or more of the following: temperature, humidity, wind speed, air pressure, and irradiance. Temperature data affects cooling and heating loads; humidity data affects human comfort and equipment heat dissipation efficiency; wind speed data affects wind power output and perceived temperature; air pressure data is related to changes in weather systems; and irradiance data affects photovoltaic output and daytime load distribution. These meteorological factors collectively constitute the exogenous driving variables affecting changes in electricity load.

[0014] Historical load data includes electricity load sequences for the target area over a historical time period, sourced from the power grid SCADA system or metering automation system. It reflects the actual electricity demand changes in the target area over a past period, including time-series characteristics such as daily cycles, weekly cycles, and seasonal trends, providing historical pattern references for load forecasting. The historical time range refers to the length of time covered by the historical data used to extract load change patterns. This historical time range is set comprehensively based on the completeness of periodic information required for load forecasting and data availability; for example, it can be set to the past 30 days. 30 days of historical data can cover complete daily cycles, weekly cycles, and load difference patterns between weekdays and holidays, providing sufficient time-series reference samples for load forecasting.

[0015] The expected energy type characteristic data includes one or more of the following: wind power connection ratio data, photovoltaic connection ratio data, and hydropower connection ratio data. It comes from the new energy power generation plan or dispatch plan system and reflects the proportion of various new energy power sources expected to be connected to the grid within a preset time window. It is used to characterize the impact of new energy output fluctuations on net load.

[0016] By collecting the above three types of data, multi-source forecast data can be formed, which can provide comprehensive input information for subsequent load forecast complexity assessment and power load forecasting. This enables the forecasting model to simultaneously perceive meteorological change patterns, historical load trends, and disturbances from new energy access, thereby improving the accuracy and adaptability of forecasts.

[0017] Furthermore, the obtained multi-source prediction data undergoes standardization and spatiotemporal alignment to generate a multidimensional input tensor. Since the multi-source prediction data originates from different data sources, each with different units of measurement and temporal resolution, it is crucial to understand that weather prediction data uses degrees Celsius for temperature and meters per second for wind speed, while historical load data uses megawatts. The temporal resolution of weather prediction data might be one hour, while historical load data might be 15 minutes. Directly inputting data with different units into the model would lead to larger variables dominating model training, thus affecting prediction accuracy.

[0018] Therefore, standardization and spatiotemporal alignment are required, recombining all data along the time and spatial dimensions to form a multidimensional input tensor. This multidimensional input tensor serves as the input to the subsequent load forecasting model, enabling the model to simultaneously utilize information from the time, space, and feature dimensions for prediction.

[0019] Specifically, the multi-source prediction data is standardized and spatiotemporally aligned to generate a multidimensional input tensor, including: A sliding window filtering method is used to eliminate random noise introduced during the meteorological forecast data collection process, and the K-nearest neighbor interpolation method is used to fill in missing items in historical load data. The Z-score standardization method is used to eliminate the dimensional differences between different variables to obtain standard feature data. The data with different time resolutions and spatial granularities in the standard feature data are unified to the standard spatiotemporal coordinate system to ensure that all data have a corresponding relationship in the time and spatial dimensions. The data are then recombined according to the time and spatial dimensions to form a multidimensional input tensor.

[0020] First, meteorological forecast data may contain high-frequency random fluctuations during acquisition and numerical weather prediction model output. This noise does not reflect the true trend of meteorological changes, and directly using it for load forecasting would introduce interference. Therefore, a sliding window filtering method is needed to eliminate the random noise introduced during the acquisition of meteorological forecast data. Sliding window filtering is a time-domain smoothing technique. Its basic principle is to set a fixed-length sliding window on the time series, statistically aggregate the data points covered by the window, and replace the original values ​​of the window's center or end points with the aggregated results, thereby filtering out high-frequency random fluctuation components.

[0021] For meteorological forecast data, the specific steps of this method are as follows: First, determine the window length, which can be set according to the fluctuation characteristics of meteorological factors, for example, set to 3; then, slide the window sequentially along the time axis, moving one time step at a time; at each window position, calculate the arithmetic mean of all data points within the window, and output this average as the smoothed value corresponding to the center position of the window. Taking a temperature series as an example, suppose the original temperature series is T1, T2, T3, T4, T5, ..., when the window length is 3, the first window covers T1, T2, T3, and calculates its average as the smoothed value of T2; the second window covers T2, T3, T4, and calculates its average as the smoothed value of T3; and so on, traversing the entire time series to obtain the smoothed temperature series.

[0022] Through the above operations, short-term random fluctuations in meteorological forecast data can be effectively suppressed, thus retaining only the low-frequency components that reflect the true trend of meteorological changes.

[0023] Secondly, missing values ​​in the historical load data are filled using the K-nearest neighbor interpolation method. Historical load data may contain missing values ​​due to sensor malfunctions or communication interruptions. K-nearest neighbor interpolation is a method for filling missing values ​​based on sample similarity. Its core idea is to find the K known samples that are most similar in features to the missing sample, and then use the values ​​of these known samples to estimate the missing value.

[0024] Specifically, for historical load data, the specific steps of this completion method are as follows: First, determine the feature space used to measure sample similarity. This can be time features such as time of day, day of the week, or holiday identifiers; meteorological features such as temperature and humidity; or load pattern features such as load values ​​from previous times. Then, set the number of nearest neighbors K. This K value can be set according to the size and distribution characteristics of the dataset, for example, set to 5. For times with missing values, search the historical data for the K known load samples most similar to the features of that time. The similarity can be measured using Euclidean distance or cosine similarity. Normalize the calculated similarity value to the range of 0 to 1. Finally, perform a weighted average of the load values ​​of the K nearest neighbor samples. The weights are determined based on the normalized similarity; for example, the weight equals the similarity of the sample divided by the sum of the similarities of the K samples.

[0025] For example, the similarities of three nearest neighbor samples are 0.9, 0.6, and 0.5, respectively, and their normalized weights are 0.45, 0.3, and 0.25, respectively. The load values ​​of the three samples are weighted and summed according to the above weights to obtain the imputed value of the missing point. Through the above operation, the integrity and continuity of historical load data are restored, which can provide a complete time series input for subsequent load forecasting.

[0026] Furthermore, the Z-score standardization method is used to eliminate the dimensional differences between different variables and obtain standard feature data. Z-score standardization is a linear transformation method based on the statistical characteristics of data. Its basic principle is to assume that the original data follows a normal distribution. By subtracting the sample mean and dividing by the sample standard deviation, the original data is transformed into standard normal distribution data with a mean of zero and a standard deviation of one, thereby eliminating the differences between different dimensions and orders of magnitude.

[0027] Specifically, for the power load forecasting scenario, the specific steps of this method are as follows: First, calculate the mean and standard deviation of each feature variable on the training set, where the mean reflects the central tendency of the feature and the standard deviation reflects the dispersion of the feature; then, apply a standardization transformation to each feature variable, subtracting the mean of the feature from the original value and dividing by the standard deviation of the feature to obtain the standardized feature value. For example, taking temperature data as an example, if the mean temperature on the training set is 15 degrees Celsius and the standard deviation is 10 degrees Celsius, then the standardized value of the original temperature of 25 degrees Celsius is 25 minus 15 divided by 10 equals 1; the standardized value of the original temperature of 5 degrees Celsius is 5 minus 15 divided by 10 equals -1.

[0028] Through the above operations, variables such as temperature, wind speed, and load, which have different dimensions, are all transformed to the same scale, making variables with different dimensions comparable and preventing variables with larger values ​​from dominating the loss function. In particular, the mean and standard deviation used in the standardization process are calculated only on the training set, and the standardization transformations used for the validation and test sets are saved to avoid data leakage.

[0029] Furthermore, data with different temporal resolutions and spatial granularities in the standard feature data are unified to a standard spatiotemporal coordinate system, ensuring that all data have a correspondence in the temporal and spatial dimensions. Spatiotemporal alignment is a key step in solving the inconsistency of multi-source heterogeneous data. Its purpose is to map data from different data sources, with different temporal sampling intervals and different spatial representations, onto a unified temporal and spatial grid, so that each data source corresponds one-to-one in the same temporal and spatial location, thereby constructing a spatiotemporally consistent multidimensional input tensor.

[0030] Specifically, the operational steps for time dimension alignment are as follows: First, determine the standard time resolution, for example, set it to 15 minutes. This resolution should not be lower than an integer approximation of the highest time resolution among all data sources. For data sources with a time resolution lower than the standard resolution, use a linear interpolation method for upsampling. Taking meteorological forecast data as an example, if the original resolution is 1 hour (one data point every 60 minutes) and the target resolution is 15 minutes, then three interpolation points are inserted at equal intervals between two adjacent original data points. The values ​​of the interpolation points are obtained by linear calculation based on the original data points at both ends. For data sources with a time resolution higher than the standard resolution, use a downsampling method for aggregation. For example, if the original resolution of historical load data is 5 minutes and the target resolution is 15 minutes, then the average of every three consecutive data points is used as the load value at a 15-minute granularity.

[0031] The specific steps for spatial dimension alignment are as follows: First, determine the standard spatial grid, such as the substation power supply area or kilometer-level grid on which load forecasting is based, as the target spatial unit. For meteorological data existing in grid form, spatial interpolation is used to map the grid data to the target spatial unit. Taking inverse distance weighted interpolation as an example, for each target spatial unit, select the meteorological grid point closest to the center of the unit, and use the square of the inverse distance as the weight to perform a weighted average of meteorological elements, with closer grid points having greater weights. Through the above interpolation, each target spatial unit can obtain corresponding meteorological characteristic values ​​such as temperature, humidity, and wind speed. For historical load data, which is itself based on substations or grids, no spatial interpolation is needed; it can be directly mapped to the target spatial unit.

[0032] After aligning the time and space dimensions, all data sources have corresponding values ​​at the same point in time and in the same spatial unit, ensuring the correspondence between the data in the time and space dimensions.

[0033] Finally, the data is recombinated along the time and spatial dimensions to form a multidimensional input tensor. The standardized and spatiotemporally aligned data is organized according to three dimensions: time step, number of spatial grids, and feature dimension, forming a three-dimensional tensor. For example, the tensor dimension could be T×N×F, where T is the time step, N is the number of spatial grids, and F is the feature dimension of each grid point. This multidimensional input tensor can be directly input into a deep learning model, enabling the model to simultaneously utilize information from the time, space, and feature dimensions for load prediction.

[0034] S20: Based on the multi-source forecast data, perform load forecast complexity assessment to determine the load forecast complexity index, call the adaptive load forecaster according to the load forecast complexity index, and predict the power load forecast result within the preset time window according to the multi-dimensional input tensor. Furthermore, based on the aforementioned multi-source forecast data, a load forecasting complexity assessment was conducted to determine the load forecasting complexity index. The load forecasting complexity index is a comprehensive indicator that quantifies the difficulty of load forecasting within a predetermined future time window, reflecting the impact of drastic changes in meteorological conditions and fluctuations in renewable energy output on load forecasting accuracy.

[0035] The load forecasting complexity index is calculated by comprehensively considering two core dimensions: first, the meteorological abrupt change index, which characterizes the drastic changes in meteorological factors such as temperature, humidity, and wind speed within a preset time window; the more drastic the changes in meteorological factors, the higher the complexity of load forecasting. Second, the renewable energy output fluctuation index, which characterizes the fluctuations in renewable energy output such as wind power and photovoltaics within a preset time window; the greater the fluctuations in renewable energy output, the higher the complexity of net load forecasting. By weighted and fused together these two dimensions, the load forecasting complexity index is obtained. A higher load forecasting complexity index indicates greater difficulty in load forecasting in the future period.

[0036] Specifically, the load forecasting complexity index is determined by assessing the load forecasting complexity based on the multi-source forecasting data, including: The meteorological mutation index is calculated based on meteorological forecast data. Specifically, the coefficient of variation of the numerical sequence of each meteorological factor within the preset time window is calculated, and the coefficient of variation of multiple meteorological factors is weighted and summed to obtain the meteorological mutation index. The ratio of the coefficient of variation of each meteorological factor to the preset coefficient of variation of the benchmark meteorological factor is used as the influence weight of the meteorological factor to obtain the influence weight of multiple meteorological factors. Based on the characteristic data of the expected access energy type, the output value sequence of each energy type within the preset time window is obtained, and the coefficient of variation of each output value sequence is calculated as the energy fluctuation index to obtain multiple energy fluctuation indices. Using the influence weight of meteorological factors corresponding to energy types as compensation coefficients, multiple energy fluctuation indices are compensated and adjusted respectively, and the adjusted multiple energy fluctuation indices are weighted and summed to obtain the new energy output fluctuation index. The load forecasting complexity index is obtained by weighting the meteorological mutation index and the new energy output fluctuation index.

[0037] First, the meteorological abrupt change index is calculated based on meteorological forecast data. The meteorological abrupt change index is used to characterize the degree of fluctuation of meteorological elements such as temperature, humidity, wind speed, and irradiance within a preset time window.

[0038] Specifically, firstly, for each meteorological factor, its numerical sequence within a preset time window is extracted, and the ratio of the standard deviation to the mean of the sequence is calculated to obtain the coefficient of variation for that meteorological factor. A larger coefficient of variation indicates more drastic fluctuations in the meteorological factor within the time window. Secondly, the coefficients of variation of multiple meteorological factors are weighted and summed to obtain the meteorological abrupt change index. The weighting coefficients can be set according to the sensitivity of each meteorological factor to load changes. For example, temperature has a significant impact on cooling and heating loads, so a temperature weighting coefficient of 0.4 can be set; irradiance has a significant impact on photovoltaic output and daytime load distribution, so an irradiance weighting coefficient of 0.3 can be set; humidity is related to perceived temperature and equipment heat dissipation efficiency, so a humidity weighting coefficient of 0.2 can be set; wind speed has a certain impact on wind power output and perceived temperature, so a wind speed weighting coefficient of 0.1 can be set. The sum of the weighting coefficients of all meteorological factors is 1. Through the above weighting allocation, the meteorological abrupt change index can accurately reflect the differentiated impact of different meteorological factors on load changes.

[0039] Secondly, the ratio of the coefficient of variation of each meteorological factor to the preset benchmark coefficient of variation of the meteorological factor is used as the influence weight of the meteorological factor, thus obtaining the influence weight of multiple meteorological factors. Specifically, the preset benchmark coefficient of variation of the meteorological factor is a reference value statistically obtained under standard meteorological conditions, reflecting the normal fluctuation level of the meteorological factor during stable periods.

[0040] When the ratio of the coefficient of variation of each meteorological factor to the coefficient of variation of the benchmark meteorological factor is greater than 1, it indicates that the fluctuation of that factor is higher than the normal level, and its influence weight on load forecasting should be increased accordingly; when the ratio is less than 1, it indicates that the fluctuation is lower than the normal level, and its influence weight should be reduced accordingly. By using this ratio as the influence weight of meteorological factors, the relative importance of each meteorological factor to load forecasting in the current period can be dynamically quantified.

[0041] Furthermore, based on the characteristic data of the expected energy types to be connected, the output value sequences of each energy type within a preset time window are obtained, and the coefficient of variation of each output value sequence is calculated as an energy fluctuation index, resulting in multiple energy fluctuation indices. Among them, the expected energy types to be connected include wind power, photovoltaic, hydropower, etc., and their output value sequences reflect the predicted power generation of various new energy sources within the preset time window.

[0042] Specifically, the output value sequence is obtained as follows: From the characteristic data of the expected energy type, the predicted wind power output, photovoltaic power output, and hydropower output corresponding to each time step within a preset time window are extracted, forming wind power output value sequences, photovoltaic power output value sequences, and hydropower output value sequences, respectively. For example, if the preset time window is the next 24 hours and the time step is 15 minutes, then each output value sequence contains predicted output values ​​for 96 time steps. For each energy type, the coefficient of variation of its output value sequence is calculated. The specific calculation steps are: first, calculate the arithmetic mean of all output values ​​in the sequence; then, calculate the sum of squares of the deviations of each output value from the mean, divide by the sequence length, and take the square root to obtain the standard deviation; finally, divide the standard deviation by the mean to obtain the coefficient of variation.

[0043] For each energy type, the coefficient of variation of its output value sequence is calculated. The larger the coefficient of variation, the more drastic the fluctuation in the energy output, and the greater the impact on net load forecasting. Taking photovoltaic (PV) output as an example, if cloud cover changes drastically during the day, the coefficient of variation of the PV output sequence will be significantly higher than that under clear weather conditions, indicating that the PV output fluctuates more on that day, and the complexity of net load forecasting increases accordingly.

[0044] Furthermore, using the influence weights of meteorological factors corresponding to energy types as compensation coefficients, multiple energy fluctuation indices are adjusted and compensated. The adjusted energy fluctuation indices are then weighted and summed to obtain the new energy output fluctuation index. Different energy types have varying degrees of dependence on meteorological conditions; for example, photovoltaic output is mainly affected by irradiance, while wind power output is mainly affected by wind speed. Therefore, the influence weights of the dominant meteorological factors corresponding to each energy type are used as compensation coefficients to adjust the fluctuation index of that energy type.

[0045] For example, multiplying the photovoltaic power output fluctuation index by the influence weight of irradiance yields the adjusted photovoltaic energy fluctuation index; multiplying the wind power output fluctuation index by the influence weight of wind speed yields the adjusted wind power energy fluctuation index; and multiplying the hydropower output fluctuation index by the influence weight of humidity yields the adjusted hydropower energy fluctuation index. After compensation and adjustment, the multiple adjusted energy fluctuation indices are weighted and summed to obtain the new energy power output fluctuation index. The weighting coefficients for the summation are set according to the proportion of each type of new energy in the total installed capacity of the target area. For example, the weighting coefficient for the photovoltaic energy fluctuation index can be set to 0.5, the wind power energy fluctuation index to 0.3, and the hydropower energy fluctuation index to 0.2. The sum of all weighting coefficients is 1. The final calculated new energy power output fluctuation index is used to comprehensively reflect the contribution of new energy power output fluctuations to the complexity of load forecasting.

[0046] Finally, a load forecasting complexity index is obtained based on a weighted evaluation of the meteorological abrupt change index and the renewable energy output fluctuation index. Specifically, the meteorological abrupt change index and the renewable energy output fluctuation index are multiplied by their respective weighting coefficients and then summed to obtain the load forecasting complexity index. The weighting coefficients can be set according to the actual application scenario. For example, during the high-temperature period in summer, the impact of meteorological abrupt changes on the load is more significant, so the weight of the meteorological abrupt change index can be appropriately increased; in areas with high renewable energy penetration, the weight of the renewable energy output fluctuation index can be appropriately increased. The higher the final calculated load forecasting complexity index, the greater the difficulty of load forecasting in the future period, requiring more forecasting units to participate in the forecasting to ensure accuracy.

[0047] In summary, through the above multi-dimensional weighted evaluation, the load forecasting complexity index can comprehensively reflect the dual impact of the severity of meteorological changes and the degree of fluctuation in new energy output on the complexity of load forecasting.

[0048] Furthermore, an adaptive load forecaster is invoked based on the load forecasting complexity index. The load forecaster is a set of forecasting models integrated from multiple load forecasting units. Each load forecasting unit is trained based on a Long Short-Term Memory (LSTM) network and has different forecasting preferences and feature extraction capabilities. The load forecaster extracts load variation patterns and meteorological driving features over time from the multidimensional input tensor. Through the integrated output of multiple load forecasting units, it generates power load forecast values ​​for each time step within a preset time window, providing accurate and reliable load forecasting results.

[0049] The construction steps of the load forecaster include: Based on the power grid operation records of the target area, several sample multi-source prediction data are collected, and the historical power load sequence of different sample multi-source prediction data within the historical time window is collected as the sample predicted power load sequence, thus obtaining several sample predicted power load sequences. The multi-source prediction data and the predicted power load sequence of the samples are used as training data, and the training data are divided into Q parts to obtain a Q sample training set; A time series forecasting model is constructed based on a long short-term memory network. The time series forecasting model is trained to convergence using the Q sample training sets, generating Q load forecasting units, which are then integrated to construct a load forecaster.

[0050] First, based on the power grid operation records of the target area, several sample multi-source prediction data are collected. Historical power load sequences within historical time windows are then collected from different sample multi-source prediction data as sample predicted power load sequences, resulting in several sample predicted power load sequences. The power grid operation records include meteorological observation data, load data, and renewable energy output data for historical time periods. The sample multi-source prediction data are input features extracted from the power grid operation records. Each sample corresponds to a historical time window, containing the meteorological factor sequence, historical load sequence, and renewable energy integration ratio within that window.

[0051] The sample predicted power load sequence is the future load true value sequence corresponding to the input sample in time, that is, the actual load value within a preset time window after the historical time window. Each input sample and its corresponding load true value sequence constitute a training sample pair.

[0052] Secondly, several samples of multi-source prediction data and several samples of predicted power load sequences are used as training data, and the training data is divided into Q equal parts to obtain a Q sample training set. Specifically, all training samples are randomly shuffled and then evenly divided into Q parts, with each sample training set containing the same number of samples. For example, Q can be set to 10, then the training data is divided into 10 equal parts. This partitioning method is used for subsequent ensemble learning, with each training set independently training a prediction unit. Each prediction unit has different prediction preferences and feature extraction capabilities due to the differences in the training data.

[0053] Furthermore, a time-series forecasting model is constructed based on a Long Short-Term Memory (LSTM) network. Each LSM is trained to convergence using Q training sets, generating Q load forecasting units, which are then integrated to construct a load forecaster. Specifically, the LSM network is a variant of a recurrent neural network suitable for time-series forecasting, capable of effectively capturing long-term dependencies and nonlinear variation patterns in load sequences. Each load forecasting unit employs the same network architecture but is trained independently using different training sets.

[0054] Optionally, the network architecture of the load prediction unit is as follows: The input layer receives a multi-dimensional input tensor, where the tensor dimension is the time step multiplied by the feature dimension. The time step refers to the number of sampling points within a historical time window; for example, if the historical window is 24 hours and the time step is 15 minutes, the time step is 96. The feature dimension refers to the total number of features included in each time step, such as meteorological factors, historical load, and the proportion of renewable energy access. Two long short-term memory (LSM) network layers are set up: the first layer contains 128 hidden units, and the second layer contains 64 hidden units. Each LSM layer is followed by a Dropout layer with a dropout rate set between 0.2 and 0.3 to suppress overfitting. The output layer is a fully connected layer that maps the hidden state of the last time step of the second LSM network layer to the load prediction values ​​for each time step within a preset time window. The number of nodes in the output layer equals the number of prediction time steps; for example, if the prediction window is 24 hours and the time step is 15 minutes, the number of nodes in the output layer is 96.

[0055] During training, key hyperparameters included a learning rate of 0.001, a training epoch count of 200, a batch size of 32, and a time step length set to the number of sampling points within the historical time window. The loss function used was mean squared error, and the optimizer was Adam. Each training unit trained independently on its corresponding training set. Training terminated when the validation set loss function value no longer decreased for several consecutive epochs; for example, if the validation set loss did not fall below the historical best value for 10 consecutive epochs, training stopped, and the model parameters were saved when the validation set loss was minimized. After training Q load forecasting units, all units were combined to form a load forecaster. During forecasting, the load forecaster could select some units to form an adapted load forecaster for prediction as needed.

[0056] Through the above integrated construction method, each load prediction unit in the load predictor has different prediction preferences due to the differences in training data. After integration, it can cover a wider feature space, thereby improving the robustness and generalization ability of prediction.

[0057] The load forecaster's invocation mechanism dynamically determines the number of units participating in the forecast based on the load forecast complexity index. When the load forecast complexity index is high, it indicates drastic future weather changes and large fluctuations in renewable energy, requiring more forecasting units to participate in the forecast to reduce forecasting errors through the averaging effect of multi-unit integration. When the load forecast complexity index is low, it indicates stable future weather and small fluctuations in renewable energy, allowing fewer forecasting units to be invoked to reduce computational overhead.

[0058] Specifically, the adapted load forecaster is invoked based on the load forecasting complexity index, and the power load forecasting results within a preset time window are obtained based on the multidimensional input tensor, including: The ratio of the load forecasting complexity index to the preset benchmark load forecasting complexity index is used as the unit selection coefficient. The ratio of the unit selection coefficient to the preset benchmark unit selection number is rounded down to obtain the number of adaptation units selected, K. The preset benchmark unit selection number is one-third of Q, and K is greater than or equal to 3 and less than or equal to Q. K prediction units are randomly selected from the Q load prediction units of the load predictor to form an adaptive load predictor. Load prediction is performed based on the multidimensional input tensor, and K power load prediction sequences are output. The K power load forecast sequences are fused by means in chronological order to obtain the power load forecast mean sequence. For the K power load prediction sequences, calculate the power load variation coefficient at the same time, and calculate the average of the K power load variation coefficients at multiple times as the prediction fluctuation coefficient; The current prediction confidence level is obtained by multiplying the ratio of the predicted volatility coefficient to the preset benchmark predicted volatility coefficient by the preset benchmark prediction confidence level and rounding down. The power load forecast mean sequence and the current forecast confidence level are used as the power load forecast result.

[0059] First, the ratio of the load forecast complexity index to the preset baseline load forecast complexity index is used as the unit selection coefficient. The preset baseline load forecast complexity index is a reference value statistically obtained under standard meteorological conditions, used to reflect the normal forecast complexity during periods of stable weather. When the ratio of the load forecast complexity index to the baseline load forecast complexity index is greater than 1, it indicates that the forecasting difficulty for the current period is higher than normal, requiring the use of more forecasting units; when the ratio is less than 1, it indicates that the forecasting difficulty is lower than normal, allowing the use of fewer forecasting units.

[0060] The ratio of the unit selection coefficient to the preset baseline unit selection number is rounded down to obtain the adaptive unit selection number K. The preset baseline unit selection number is one-third of the total number of units Q in the load forecaster. For example, when Q equals 9, the baseline unit selection number is 3. Rounding down ensures that K is an integer, and the lower limit of K is set to 3, and the upper limit is set to Q, guaranteeing that at least 3 units are called to ensure forecast stability, and that the number does not exceed the total number of units.

[0061] Secondly, K forecasting units are randomly selected from the Q load forecasting units of the load forecaster to form an adaptive load forecaster. This adaptive load forecaster performs load forecasting based on a multidimensional input tensor and outputs K power load forecast sequences. Specifically, each selected load forecasting unit independently performs forward computation on the same multidimensional input tensor, outputting the load forecast values ​​for each time step within a preset time window, forming a power load forecast sequence. Due to differences in the training data of each forecasting unit, there are certain differences between their output forecast sequences. The K forecasting units output a total of K power load forecast sequences, each containing the same number of time step forecast values.

[0062] Then, the K power load forecast sequences are fused by mean fusion in chronological order to obtain a power load forecast mean sequence. Specifically, for each time step, the arithmetic mean of the predicted values ​​of the K power load forecast sequences at that time step is calculated to obtain the fused forecast value for that time step. This process is repeated for all time steps, and the fused forecast values ​​for each time step are arranged in chronological order to form the power load forecast mean sequence. Mean fusion effectively reduces the impact of random errors from individual forecast units on the final forecast result.

[0063] Simultaneously, for K power load forecast sequences, the coefficient of variation of power load at the same time step is calculated, and the average of the K power load coefficients of variation at multiple time steps is calculated as the forecast fluctuation coefficient. The coefficient of variation is equal to the standard deviation of the K forecast values ​​at the same time step divided by the average, and is used to reflect the consistency of the forecast results of each forecasting unit at that time step. The smaller the power load coefficient of variation, the more consistent the forecast results of each forecasting unit, and the higher the reliability of the forecast results; the larger the power load coefficient of variation, the greater the forecast discrepancy between the forecasting units, and the higher the uncertainty of the forecast results. After calculating the power load coefficient of variation for all time steps, the average value is taken as the forecast fluctuation coefficient, which is used to quantify the degree of uncertainty of this forecast as a whole.

[0064] Furthermore, the current prediction confidence level is obtained by multiplying the ratio of the predicted fluctuation coefficient to the preset baseline predicted fluctuation coefficient by the preset baseline prediction confidence level and rounding down. The preset baseline predicted fluctuation coefficient is a reference value statistically obtained under standard prediction conditions, representing the normal uncertainty level of load forecasting when weather is stable and renewable energy output fluctuations are small. The preset baseline prediction confidence level is the confidence benchmark value output under standard conditions; for example, it can be set to 0.9, representing a 90% reliability of the prediction result under standard prediction conditions.

[0065] When the ratio of the predicted volatility coefficient to the preset benchmark predicted volatility coefficient is greater than 1, it indicates that the uncertainty of this prediction is higher than normal, and the current prediction confidence level decreases accordingly; when the ratio is less than 1, the current prediction confidence level increases accordingly. The current prediction confidence level is limited to between 0 and 1 by rounding down.

[0066] Finally, the power load forecast mean sequence and the current forecast confidence level are output as the power load forecast result. The power load forecast mean sequence provides the point forecast value, and the current forecast confidence level provides a quantitative indicator of the reliability of the forecast result. Together, they constitute a complete power load forecast result, providing a basis for subsequent extreme weather judgment and correction. In summary, through the above adaptive invocation mechanism, more units are invoked to ensure accuracy when the forecast difficulty is high, and fewer units are invoked to save computational resources when the forecast difficulty is low, thus achieving a balance between forecast accuracy and computational efficiency.

[0067] S30: Based on the multi-source prediction data, determine whether the target area is experiencing non-extreme or extreme weather within a preset time window. If it is non-extreme weather, output the power load prediction result. If it is extreme weather, call the extreme weather correction model to correct the power load prediction result and output the corrected load prediction result.

[0068] Furthermore, based on the aforementioned multi-source forecast data, it is determined whether the target area experiences non-extreme or extreme weather within a preset time window. Extreme weather refers to weather phenomena where meteorological factors such as temperature, humidity, wind speed, and irradiance deviate significantly from normal ranges, such as extreme heat waves, extreme cold waves, typhoons, and heavy rainstorms. Extreme weather can cause drastic fluctuations in power load, and the prediction accuracy of conventional load forecasting models decreases significantly under extreme weather conditions. Therefore, it is necessary to first determine whether extreme weather exists within the preset time window in order to decide whether to revise the forecast results.

[0069] Specifically, determining whether a target area experiences non-extreme or extreme weather within a preset time window based on the multi-source prediction data includes: The forecast sequence of multiple meteorological factors within the preset time window is extracted from meteorological forecast data. The multiple meteorological factors include at least three of temperature, humidity, wind speed, air pressure, and irradiance. The extreme values, magnitude of change, and rate of change of each meteorological factor prediction sequence are calculated separately. Based on the calculation results, multiple single-factor extreme indices of multiple meteorological factors are evaluated. The multiple single-factor extreme indices are weighted and summed to obtain a comprehensive meteorological extreme index. When the comprehensive meteorological extreme index is greater than or equal to the preset extreme weather trigger threshold, the target area is determined to be an extreme weather scenario within the preset time window; When the comprehensive meteorological extreme index is less than the extreme weather trigger threshold, the target area is determined to be a non-extreme weather scenario within the preset time window.

[0070] First, forecast sequences of multiple meteorological factors within a preset time window are extracted from meteorological forecast data. These meteorological factors include at least three of the following: temperature, humidity, wind speed, air pressure, and irradiance. The temperature forecast sequence reflects the trend of air temperature changes; extreme high or low temperatures are the main factors causing abnormal load fluctuations. The humidity forecast sequence reflects the water vapor content in the air; in high-temperature and high-humidity environments, the perceived temperature increases, which further increases the cooling load. The wind speed forecast sequence affects perceived temperature and wind power output. The air pressure forecast sequence is related to changes in weather systems; drastic fluctuations are often accompanied by extreme weather events. The irradiance forecast sequence reflects the intensity of solar radiation, affecting photovoltaic output and daytime load distribution.

[0071] Secondly, the extreme values, amplitudes of change, and rates of change of each meteorological factor's prediction sequence are calculated separately. Based on the calculation results, multiple single-factor extreme indices for multiple meteorological factors are evaluated. Specifically, extreme values ​​refer to the maximum and minimum values ​​in the sequence, reflecting the extreme degree reached by the meteorological factor within the time window. Amplitude of change refers to the difference between the maximum and minimum values ​​of the sequence, reflecting the fluctuation range of the meteorological factor within the time window. Rate of change refers to the amount of change in the sequence value per unit time, such as the hourly temperature change, reflecting the drastic degree of change in the meteorological factor.

[0072] The three indicators mentioned above are weighted and summed according to preset weights to obtain the single-factor extreme index of the meteorological factor. In the weighted summation, the weight coefficients of each indicator are set according to its ability to represent the severity of extreme weather. For example, the extreme value weight coefficient can be set to 0.5, the variation range weight coefficient to 0.3, and the change rate weight coefficient to 0.2. Extreme values ​​directly reflect the extreme degree reached by the meteorological factor and are the core indicator for judging extreme weather, therefore they are given the highest weight. Variation range reflects the fluctuation range of the meteorological factor, and change rate reflects the drastic degree of change of the meteorological factor; both are given lower weights as auxiliary indicators. Through the above weighted summation, the single-factor extreme index can comprehensively reflect the extreme degree of the meteorological factor in three dimensions: extreme value intensity, fluctuation range, and change rate.

[0073] Then, a weighted sum of multiple single-factor extreme indices is obtained to obtain a comprehensive meteorological extreme index. The weight coefficients of each meteorological factor can be set according to its sensitivity to load. For example, temperature has the greatest impact on load, so a weight coefficient of 0.4 can be set for temperature; irradiance has a significant impact on photovoltaic output and daytime load, so an irradiance weight coefficient of 0.2 can be set for irradiance; humidity, wind speed, and air pressure can each be set with weight coefficients of 0.2, 0.1, and 0.1, respectively, with the sum of the weight coefficients being 1. The comprehensive meteorological extreme index comprehensively reflects the extreme degree of multiple meteorological factors within a preset time window.

[0074] Furthermore, when the calculated comprehensive meteorological extreme index is greater than or equal to the preset extreme weather trigger threshold, the target area is determined to be in an extreme weather scenario within a preset time window. The preset extreme weather trigger threshold is a critical value obtained based on historical extreme weather events, and can be set to, for example, 0.7. When the comprehensive meteorological extreme index reaches 0.7 or higher, it indicates that the meteorological conditions have reached or exceeded the intensity of historical extreme weather events, and the area is determined to be in an extreme weather scenario.

[0075] When the comprehensive meteorological extreme index is less than the extreme weather trigger threshold, the target area is determined to be in a non-extreme weather scenario within a preset time window. At this time, the meteorological conditions are within the normal range, and the conventional load forecasting model can provide sufficient forecast accuracy, so there is no need to call the extreme weather correction model.

[0076] The multi-factor comprehensive evaluation method described above can accurately identify whether extreme weather exists within a preset time window, providing a quantitative basis for whether to subsequently call the extreme weather correction model. Specifically, this method comprehensively considers the extreme values, amplitudes, and rates of change of multiple meteorological factors, enabling a comprehensive characterization of the intensity features of extreme weather and avoiding misjudgments or omissions caused by relying on only a single meteorological factor.

[0077] If the weather is determined to be non-extreme, it indicates that the meteorological conditions will be stable in the future period, and the conventional load forecasting model can provide sufficient forecast accuracy. In this case, the power load forecasting result obtained in step S20 can be directly output without additional correction.

[0078] However, if the weather is classified as extreme, indicating the presence of extreme weather events in the future, the predictions from conventional load forecasting models may contain significant biases. Therefore, it is necessary to use an extreme weather correction model to adjust the power load forecast. The extreme weather correction model is a correction model built on a long short-term memory network. By learning the mapping relationship between meteorological factors and load deviations under historical extreme weather events, it can compensate for and adjust the conventional forecast results. The corrected load forecast results can more accurately reflect the load change patterns under extreme weather conditions, effectively reducing forecast errors under extreme weather conditions.

[0079] Specifically, the extreme weather correction model is invoked to correct the power load forecast results, and the corrected load forecast results are output, including: If it is extreme weather, the difference between the comprehensive meteorological extreme index and the extreme weather trigger threshold is calculated as the extreme index difference; Construct an extreme weather correction model, wherein the extreme weather correction model includes P extreme weather correction branches; The ratio of the extreme exponential difference to the preset benchmark extreme exponential deviation is multiplied by the initial number of branches and rounded down to obtain the number of adaptive branches L, where the initial number of branches is one-third P, and L is greater than or equal to 2 and less than or equal to P. L branches are randomly selected from the P extreme weather correction branches. Based on multiple single-factor extreme indices of multiple meteorological factors, the mean sequence of power load prediction in the power load prediction results is corrected, and L initial corrected power load prediction sequences are output. The mean is then calculated to obtain the corrected power load prediction sequence. For the L initial corrected power load prediction sequences, the corrected power load variation coefficient at the same time is calculated, and the average value of the L corrected power load variation coefficients at multiple times is calculated as the prediction correction fluctuation coefficient. The initial corrected prediction confidence level is obtained by multiplying the ratio of the predicted corrected volatility coefficient to the preset benchmark predicted corrected volatility coefficient by the preset benchmark corrected prediction confidence level and rounding down. The product of the current prediction confidence and the initial revised prediction confidence is used as the current revised prediction confidence. The corrected power load forecast sequence and the current corrected forecast confidence level are output as the corrected load forecast result.

[0080] First, if the weather is determined to be extreme, the difference between the comprehensive meteorological extreme index and the extreme weather trigger threshold is calculated as the extreme index difference. This extreme index difference reflects the extent to which the severity of the current extreme weather exceeds the trigger threshold; the larger the difference, the more severe the extreme weather, and the stronger the correction required. For example, if the comprehensive meteorological extreme index is 0.85 and the extreme weather trigger threshold is 0.7, then the extreme index difference is 0.15.

[0081] Secondly, an extreme weather correction model is constructed. This model is an ensemble of correction models built from P extreme weather correction branches. Each branch is trained using a long short-term memory network to learn the mapping relationship between meteorological factors and load forecast bias under historical extreme weather events. This allows for targeted compensation of conventional load forecast results during extreme weather events, reducing forecast errors under extreme conditions. The value of P can be set according to the training sample size and computational resources; for example, it can be set to 9.

[0082] The steps involved in constructing an extreme weather correction model include: Based on the power grid operation records of the target area, several sample single-factor extreme index sets and several sample power load prediction sequences are collected. The historical power load sequences of different sample single-factor extreme index sets and sample power load prediction sequences within the historical time window are collected as sample corrected power load prediction sequences, resulting in several sample corrected power load prediction sequences. The aforementioned sample single-factor extreme index sets, sample power load prediction sequences, and sample modified power load prediction sequences are used as training data and divided into P equal parts to obtain P modified training sets. The long short-term memory network is trained to convergence using the P corrected training sets, generating P extreme weather correction branches, which are then combined to obtain the extreme weather correction model.

[0083] First, based on the power grid operation records of the target area, several sample single-factor extreme index sets and several sample power load prediction sequences are collected. Then, the historical power load sequences of different sample single-factor extreme index sets and sample power load prediction sequences within the historical time window are collected as sample corrected power load prediction sequences, resulting in several sample corrected power load prediction sequences.

[0084] Specifically, the power grid operation record includes meteorological and load data during historical extreme weather events. The sample single-factor extreme index set is a set of single-factor extreme indices extracted from various meteorological factors during historical extreme weather events, including extreme indices for temperature, humidity, wind speed, and irradiance. The sample power load prediction sequence is the predicted load sequence output by the conventional load prediction model under the historical extreme weather event. The sample corrected power load prediction sequence is the actual load ground value sequence corresponding to the input sample in time, i.e., the corrected target value. Each training sample contains three parts: input features include the single-factor extreme index set and the conventional predicted load sequence, and the supervision label is the actual load ground value sequence.

[0085] Secondly, several sample single-factor extreme index sets, several sample power load prediction sequences, and several sample modified power load prediction sequences are used as training data, and divided into P equal parts to obtain P modified training sets. All training samples are randomly shuffled and then uniformly divided into P parts, with each modified training set containing the same number of samples. For example, P can be set to 9, dividing the training data into 9 equal parts. This partitioning method is used for subsequent ensemble learning, with each extreme weather modification branch trained independently, exhibiting different modification preferences due to differences in the training data.

[0086] Then, P modified training sets are used to train the Long Short-Term Memory Network until convergence, generating P extreme weather correction branches, which are combined to obtain the extreme weather correction model.

[0087] Optionally, the network architecture of the extreme weather correction branch is as follows: The input layer receives a corrected input feature vector, which is composed of a set of sample single-factor extreme indices and a sample power load prediction sequence. The set of sample single-factor extreme indices includes multiple scalar values ​​such as temperature extreme index, humidity extreme index, wind speed extreme index, and irradiance extreme index; the sample power load prediction sequence is a sequence of load prediction values ​​for each time step within a preset time window. The number of nodes in the input layer is equal to the number of single-factor extreme indices plus the number of prediction time steps. Two layers of Long Short-Term Memory (LSTM) networks are set up. The first layer contains 64 hidden units, and the second layer contains 32 hidden units. Each LSM network is followed by a Dropout layer with a dropout rate set between 0.2 and 0.3 to suppress overfitting. The output layer is a fully connected layer that maps the hidden state of the last time step of the second LSM network to the corrected load prediction values ​​for each time step within the preset time window. The number of nodes in the output layer is equal to the number of prediction time steps.

[0088] During training, key hyperparameters included a learning rate of 0.001, 150 training epochs, and a batch size of 16. Mean squared error was used as the loss function, and Adam was used as the optimizer. Each correction branch was trained independently on its corresponding correction training set. Training terminated when the validation set loss function value no longer decreased for several consecutive epochs; for example, if the validation set loss did not fall below the historical best value for 10 consecutive epochs, training stopped, and the model parameters at the minimum validation set loss were saved.

[0089] After training the P extreme weather correction branches, all branches are combined to form an extreme weather correction model. During correction, the extreme weather correction model can select some branches to form an adaptive correction model based on the magnitude of the extreme index difference. Through this integrated construction method, each extreme weather correction branch has different correction preferences due to the differences in training data. After integration, it can cover a wider range of extreme weather scenario feature spaces, thereby improving the robustness and generalization ability of the correction.

[0090] Furthermore, the ratio of the extreme index difference to the preset baseline extreme index deviation is multiplied by the initial number of branches and rounded down to obtain the number of adaptive branches, L. The preset baseline extreme index deviation is a reference deviation value under standard extreme weather conditions; for example, it can be set to 0.1, representing the typical extent to which the comprehensive meteorological extreme index exceeds the trigger threshold under a standard extreme weather event. The initial number of branches is one-third of P; for example, when P equals 9, the initial number of branches is 3. When the ratio of the extreme index difference to the baseline extreme index deviation is greater than 1, it indicates that the severity of the extreme weather is higher than the standard level, requiring more correction branches to be invoked; when the ratio is less than 1, fewer correction branches are invoked.

[0091] Meanwhile, the number of adaptation branches L is set to a lower limit of 2 and an upper limit of P to ensure that at least 2 branches are called to guarantee the stability of the correction, and does not exceed the total number of branches.

[0092] Furthermore, L branches are randomly selected from the P extreme weather correction branches. Based on multiple single-factor extreme indices of multiple meteorological factors, the mean sequence of power load forecasts in the power load forecast results is corrected, outputting L initial corrected power load forecast sequences. Specifically, each selected extreme weather correction branch receives a single-factor extreme index vector and the mean sequence of power load forecasts as input, and outputs a corrected load forecast sequence. Due to differences in the training data of each correction branch, there are certain differences between the corrected sequences output by each branch. The mean of the L initial corrected power load forecast sequences is calculated at each time step to obtain the corrected power load forecast sequence.

[0093] Simultaneously, for L initial corrected power load forecast sequences, the coefficient of variation of corrected power load at the same time point is calculated, and the average of the L corrected power load coefficients of variation at multiple time points is calculated as the prediction correction fluctuation coefficient. The corrected power load coefficient of variation is equal to the standard deviation of the L initial corrected power load forecast values ​​at the same time point divided by the average value, and is used to reflect the consistency of the correction results of each extreme weather correction branch at that time point. The smaller the corrected power load coefficient of variation, the more consistent the correction results of each extreme weather correction branch are, and the higher the reliability of the correction results. The average of the corrected power load coefficients of variation calculated for all time steps is taken as the prediction correction fluctuation coefficient. This prediction correction fluctuation coefficient is used to characterize the degree of uncertainty of the correction of load forecast results by each extreme weather correction branch under extreme weather scenarios. The smaller the coefficient, the more similar the correction results of each correction branch are, and the higher the reliability of the correction results; the larger the coefficient, the greater the divergence of the correction results between the correction branches, and the higher the uncertainty of the correction results.

[0094] Furthermore, the initial corrected prediction confidence level is obtained by multiplying the ratio of the predicted correction fluctuation coefficient to the preset baseline predicted correction fluctuation coefficient by the preset baseline corrected prediction confidence level and rounding down. Here, the preset baseline predicted correction fluctuation coefficient is a reference value statistically obtained under standard extreme weather conditions, and the preset baseline corrected prediction confidence level is the confidence level benchmark value output under standard conditions, which can be set to, for example, 0.85.

[0095] When the ratio of the predicted correction volatility coefficient to the benchmark predicted correction volatility coefficient is greater than 1, it indicates that the uncertainty of this correction is higher than the normal level, and the confidence level of the initial correction prediction decreases accordingly; when the ratio is less than 1, the confidence level of the initial correction prediction increases accordingly.

[0096] Furthermore, the product of the current prediction confidence and the initial revised prediction confidence is taken as the current revised prediction confidence. Here, the current prediction confidence is the regular prediction confidence obtained in step S20, and the initial revised prediction confidence is the confidence of the revision process itself. Multiplying the two together reflects the overall confidence level from the regular prediction to the revised prediction.

[0097] Finally, the corrected power load forecast sequence and the current corrected forecast confidence level are output as the corrected load forecast result. In summary, through the above-mentioned extreme weather correction mechanism, targeted compensation is provided for conventional forecast results under extreme weather scenarios, effectively reducing forecast errors under extreme weather conditions; at the same time, through multi-branch integration and confidence quantification, the reliability assessment of the correction result is provided, offering more comprehensive information support for dispatching decisions.

[0098] In summary, the embodiments of this application have at least the following technical effects: This invention first collects multi-source forecast data, standardizes and aligns it spatiotemporally to generate a multi-dimensional input tensor, providing a unified and high-quality data foundation for subsequent forecasts. Second, it assesses the complexity of load forecasting based on the multi-source forecast data, determines a load forecasting complexity index, and adaptively calls an appropriate load forecaster based on this index. This achieves dynamic matching between the complexity of the forecasting model and the complexity of meteorological changes, avoiding the problems of insufficient accuracy of a single model under complex weather conditions and computational redundancy under stable weather conditions. Third, it determines whether the weather is extreme based on the multi-source forecast data. If it is not extreme, the forecast result is output directly; if it is extreme, an extreme weather correction model is called for correction. Through a targeted compensation mechanism for extreme scenarios, the forecast accuracy under extreme weather conditions is improved.

[0099] Ultimately, this invention solves the problems in the prior art of difficulty in adaptively adjusting forecasting strategies according to meteorological complexity and the lack of effective correction mechanisms in extreme weather scenarios, and achieves a balanced optimization of load forecasting accuracy and computational efficiency.

[0100] Example 2, as Figure 2 As shown, based on the same inventive concept as the intelligent power load forecasting method based on meteorological assessment provided in Embodiment 1, this embodiment of the invention also provides an intelligent power load forecasting system based on meteorological assessment, comprising: The data acquisition and processing module 11 is used to acquire multi-source prediction data of the target area within a preset time window, perform standardization processing and spatiotemporal alignment on the multi-source prediction data, and generate a multi-dimensional input tensor. The load forecasting module 12 is used to assess the load forecasting complexity based on the multi-source forecasting data, determine the load forecasting complexity index, call the adaptive load forecaster based on the load forecasting complexity index, and predict the power load forecasting result within a preset time window based on the multi-dimensional input tensor. The weather determination and correction module 13 is used to determine whether the target area is experiencing non-extreme or extreme weather within a preset time window based on the multi-source prediction data. If it is non-extreme weather, the module outputs the power load prediction result. If it is extreme weather, the module calls the extreme weather correction model to correct the power load prediction result and outputs the corrected load prediction result.

[0101] The data acquisition and processing module 11 is specifically used for: Collect multi-source prediction data of the target area within a preset time window, standardize and spatiotemporally align the multi-source prediction data, and generate a multi-dimensional input tensor.

[0102] The multi-source forecast data includes meteorological forecast data, historical load data, and expected energy type characteristic data. The meteorological forecast data includes one or more of temperature data, humidity data, wind speed data, air pressure data, and irradiance data. The historical load data includes the electricity load sequence of the target area within a historical time range. The expected energy type characteristic data includes one or more of wind power access ratio data, photovoltaic access ratio data, and hydropower access ratio data.

[0103] Furthermore, the multi-source prediction data is standardized and spatiotemporally aligned to generate a multidimensional input tensor, including: A sliding window filtering method is used to eliminate random noise introduced during the meteorological forecast data collection process, and the K-nearest neighbor interpolation method is used to fill in missing items in historical load data. The Z-score standardization method is used to eliminate the dimensional differences between different variables to obtain standard feature data. The data with different time resolutions and spatial granularities in the standard feature data are unified to the standard spatiotemporal coordinate system to ensure that all data have a corresponding relationship in the time and spatial dimensions. The data are then recombined according to the time and spatial dimensions to form a multidimensional input tensor.

[0104] Specifically, the load forecasting module 12 is used for: The load forecasting complexity index is determined based on the multi-source forecasting data, including: The meteorological mutation index is calculated based on meteorological forecast data. Specifically, the coefficient of variation of the numerical sequence of each meteorological factor within the preset time window is calculated, and the coefficient of variation of multiple meteorological factors is weighted and summed to obtain the meteorological mutation index. The ratio of the coefficient of variation of each meteorological factor to the preset coefficient of variation of the benchmark meteorological factor is used as the influence weight of the meteorological factor to obtain the influence weight of multiple meteorological factors. Based on the characteristic data of the expected access energy type, the output value sequence of each energy type within the preset time window is obtained, and the coefficient of variation of each output value sequence is calculated as the energy fluctuation index to obtain multiple energy fluctuation indices. Using the influence weight of meteorological factors corresponding to energy types as compensation coefficients, multiple energy fluctuation indices are compensated and adjusted respectively, and the adjusted multiple energy fluctuation indices are weighted and summed to obtain the new energy output fluctuation index. The load forecasting complexity index is obtained by weighting the meteorological mutation index and the new energy output fluctuation index.

[0105] The construction steps of the load forecaster include: Based on the power grid operation records of the target area, several sample multi-source prediction data are collected, and the historical power load sequence of different sample multi-source prediction data within the historical time window is collected as the sample predicted power load sequence, thus obtaining several sample predicted power load sequences. The multi-source prediction data and the predicted power load sequence of the samples are used as training data, and the training data are divided into Q parts to obtain a Q sample training set; A time series forecasting model is constructed based on a long short-term memory network. The time series forecasting model is trained to convergence using the Q sample training sets, generating Q load forecasting units, which are then integrated to construct a load forecaster.

[0106] The load forecasting complexity index is used to call the adaptive load forecaster, and the power load forecasting results within a preset time window are obtained based on the multidimensional input tensor prediction, including: The ratio of the load forecasting complexity index to the preset benchmark load forecasting complexity index is used as the unit selection coefficient. The ratio of the unit selection coefficient to the preset benchmark unit selection number is rounded down to obtain the number of adaptation units selected, K. The preset benchmark unit selection number is one-third of Q, and K is greater than or equal to 3 and less than or equal to Q. K prediction units are randomly selected from the Q load prediction units of the load predictor to form an adaptive load predictor. Load prediction is performed based on the multidimensional input tensor, and K power load prediction sequences are output. The K power load forecast sequences are fused by means in chronological order to obtain the power load forecast mean sequence. For the K power load prediction sequences, calculate the power load variation coefficient at the same time, and calculate the average of the K power load variation coefficients at multiple times as the prediction fluctuation coefficient; The current prediction confidence level is obtained by multiplying the ratio of the predicted volatility coefficient to the preset benchmark predicted volatility coefficient by the preset benchmark prediction confidence level and rounding down. The power load forecast mean sequence and the current forecast confidence level are used as the power load forecast result.

[0107] Specifically, the weather determination and correction module 13 is used for: Determining whether a target area experiences non-extreme or extreme weather within a preset time window based on the multi-source prediction data includes: The forecast sequence of multiple meteorological factors within the preset time window is extracted from meteorological forecast data. The multiple meteorological factors include at least three of temperature, humidity, wind speed, air pressure, and irradiance. The extreme values, magnitude of change, and rate of change of each meteorological factor prediction sequence are calculated separately. Based on the calculation results, multiple single-factor extreme indices of multiple meteorological factors are evaluated. The multiple single-factor extreme indices are weighted and summed to obtain a comprehensive meteorological extreme index. When the comprehensive meteorological extreme index is greater than or equal to the preset extreme weather trigger threshold, the target area is determined to be an extreme weather scenario within the preset time window; When the comprehensive meteorological extreme index is less than the extreme weather trigger threshold, the target area is determined to be a non-extreme weather scenario within the preset time window.

[0108] Furthermore, the extreme weather correction model is invoked to correct the power load forecast results, and the corrected load forecast results are output, including: If it is extreme weather, the difference between the comprehensive meteorological extreme index and the extreme weather trigger threshold is calculated as the extreme index difference; Construct an extreme weather correction model, wherein the extreme weather correction model includes P extreme weather correction branches; The ratio of the extreme exponential difference to the preset benchmark extreme exponential deviation is multiplied by the initial number of branches and rounded down to obtain the number of adaptive branches L, where the initial number of branches is one-third P, and L is greater than or equal to 2 and less than or equal to P. L branches are randomly selected from the P extreme weather correction branches. Based on multiple single-factor extreme indices of multiple meteorological factors, the mean sequence of power load prediction in the power load prediction results is corrected, and L initial corrected power load prediction sequences are output. The mean is then calculated to obtain the corrected power load prediction sequence. For the L initial corrected power load prediction sequences, the corrected power load variation coefficient at the same time is calculated, and the average value of the L corrected power load variation coefficients at multiple times is calculated as the prediction correction fluctuation coefficient. The initial corrected prediction confidence level is obtained by multiplying the ratio of the predicted corrected volatility coefficient to the preset benchmark predicted corrected volatility coefficient by the preset benchmark corrected prediction confidence level and rounding down. The product of the current prediction confidence and the initial revised prediction confidence is used as the current revised prediction confidence. The corrected power load forecast sequence and the current corrected forecast confidence level are output as the corrected load forecast result.

[0109] The steps involved in constructing an extreme weather correction model include: Based on the power grid operation records of the target area, several sample single-factor extreme index sets and several sample power load prediction sequences are collected. The historical power load sequences of different sample single-factor extreme index sets and sample power load prediction sequences within the historical time window are collected as sample corrected power load prediction sequences, resulting in several sample corrected power load prediction sequences. The aforementioned sample single-factor extreme index sets, sample power load prediction sequences, and sample modified power load prediction sequences are used as training data and divided into P equal parts to obtain P modified training sets. The long short-term memory network is trained to convergence using the P corrected training sets, generating P extreme weather correction branches, which are then combined to obtain the extreme weather correction model.

Claims

1. A method for intelligent power load forecasting based on meteorological assessment, characterized in that, include: Collect multi-source prediction data of the target area within a preset time window, and perform standardization and spatiotemporal alignment on the multi-source prediction data to generate a multi-dimensional input tensor; The load forecasting complexity index is determined by evaluating the load forecasting complexity based on the multi-source forecasting data. The load forecasting complexity index is then used to call the adaptive load forecaster. The power load forecasting result within the preset time window is obtained by predicting the multi-dimensional input tensor. The meteorological abrupt change index is calculated based on the meteorological forecasting data. The load forecasting complexity index is obtained by weighted evaluation based on the meteorological abrupt change index and the new energy output fluctuation index. Based on the multi-source prediction data, it is determined whether the target area will experience non-extreme or extreme weather within a preset time window. If it is non-extreme weather, the power load prediction result is output; if it is extreme weather, the extreme weather correction model is called to correct the power load prediction result, and the corrected load prediction result is output, including: If it is extreme weather, the difference between the comprehensive meteorological extreme index and the extreme weather trigger threshold is calculated as the extreme index difference; Construct an extreme weather correction model, wherein the extreme weather correction model includes P extreme weather correction branches; The ratio of the extreme exponential difference to the preset benchmark extreme exponential deviation is multiplied by the initial number of branches and rounded down to obtain the number of adaptive branches L, where the initial number of branches is one-third P, and L is greater than or equal to 2 and less than or equal to P. L branches are randomly selected from the P extreme weather correction branches. Based on multiple single-factor extreme indices of multiple meteorological factors, the mean sequence of power load prediction in the power load prediction results is corrected, and L initial corrected power load prediction sequences are output. The mean is then calculated to obtain the corrected power load prediction sequence. For the L initial corrected power load prediction sequences, the corrected power load variation coefficient at the same time is calculated, and the average value of the L corrected power load variation coefficients at multiple times is calculated as the prediction correction fluctuation coefficient. The initial corrected prediction confidence level is obtained by multiplying the ratio of the predicted corrected volatility coefficient to the preset benchmark predicted corrected volatility coefficient by the preset benchmark corrected prediction confidence level and rounding down. The product of the current prediction confidence and the initial revised prediction confidence is used as the current revised prediction confidence. The corrected power load forecast sequence and the current corrected forecast confidence level are output as the corrected load forecast result.

2. The intelligent power load forecasting method based on meteorological assessment according to claim 1, characterized in that, The multi-source forecast data includes meteorological forecast data, historical load data, and expected energy type characteristic data. The meteorological forecast data includes one or more of temperature data, humidity data, wind speed data, air pressure data, and irradiance data. The historical load data includes the electricity load sequence of the target area within a historical time range. The expected energy type characteristic data includes one or more of wind power access ratio data, photovoltaic access ratio data, and hydropower access ratio data.

3. The intelligent power load forecasting method based on meteorological assessment according to claim 1, characterized in that, The multi-source prediction data is standardized and spatiotemporally aligned to generate a multidimensional input tensor, including: A sliding window filtering method is used to eliminate random noise introduced during the meteorological forecast data collection process, and the K-nearest neighbor interpolation method is used to fill in missing items in historical load data. The Z-score standardization method is used to eliminate the dimensional differences between different variables to obtain standard feature data. The data with different time resolutions and spatial granularities in the standard feature data are unified to the standard spatiotemporal coordinate system to ensure that all data have a corresponding relationship in the time and spatial dimensions. The data are then recombined according to the time and spatial dimensions to form a multidimensional input tensor.

4. The intelligent power load forecasting method based on meteorological assessment according to claim 1, characterized in that, Based on the multi-source forecasting data, a load forecasting complexity index is determined through load forecasting complexity assessment, including: The meteorological mutation index is calculated based on meteorological forecast data. Specifically, the coefficient of variation of the numerical sequence of each meteorological factor within the preset time window is calculated, and the coefficient of variation of multiple meteorological factors is weighted and summed to obtain the meteorological mutation index. The ratio of the coefficient of variation of each meteorological factor to the preset coefficient of variation of the benchmark meteorological factor is used as the influence weight of the meteorological factor to obtain the influence weight of multiple meteorological factors. Based on the characteristic data of the expected access energy type, the output value sequence of each energy type within the preset time window is obtained, and the coefficient of variation of each output value sequence is calculated as the energy fluctuation index to obtain multiple energy fluctuation indices. Using the influence weight of meteorological factors corresponding to energy types as compensation coefficients, multiple energy fluctuation indices are compensated and adjusted respectively, and the adjusted multiple energy fluctuation indices are weighted and summed to obtain the new energy output fluctuation index. The load forecasting complexity index is obtained by weighting the meteorological mutation index and the new energy output fluctuation index.

5. The intelligent power load forecasting method based on meteorological assessment according to claim 1, characterized in that, The steps involved in building a load forecaster include: Based on the power grid operation records of the target area, several sample multi-source prediction data are collected, and the historical power load sequence of different sample multi-source prediction data within the historical time window is collected as the sample predicted power load sequence, thus obtaining several sample predicted power load sequences. The multi-source prediction data and the predicted power load sequence of the samples are used as training data, and the training data are divided into Q parts to obtain a Q sample training set; A time series forecasting model is constructed based on a long short-term memory network. The time series forecasting model is trained to convergence using the Q sample training sets, generating Q load forecasting units, which are then integrated to construct a load forecaster.

6. The intelligent power load forecasting method based on meteorological assessment according to claim 5, characterized in that, Based on the load forecasting complexity index, the adaptive load forecaster is invoked, and the power load forecasting results within a preset time window are obtained based on the multidimensional input tensor, including: The ratio of the load forecasting complexity index to the preset benchmark load forecasting complexity index is used as the unit selection coefficient. The ratio of the unit selection coefficient to the preset benchmark unit selection number is rounded down to obtain the number of adaptation units selected, K. The preset benchmark unit selection number is one-third of Q, and K is greater than or equal to 3 and less than or equal to Q. K prediction units are randomly selected from the Q load prediction units of the load predictor to form an adaptive load predictor. Load prediction is performed based on the multidimensional input tensor, and K power load prediction sequences are output. The K power load forecast sequences are fused by means in chronological order to obtain the power load forecast mean sequence. For the K power load prediction sequences, calculate the power load variation coefficient at the same time, and calculate the average of the K power load variation coefficients at multiple times as the prediction fluctuation coefficient; The current prediction confidence level is obtained by multiplying the ratio of the predicted volatility coefficient to the preset benchmark predicted volatility coefficient by the preset benchmark prediction confidence level and rounding down. The power load forecast mean sequence and the current forecast confidence level are used as the power load forecast result.

7. The intelligent power load forecasting method based on meteorological assessment according to claim 1, characterized in that, Determining whether a target area experiences non-extreme or extreme weather within a preset time window based on the multi-source prediction data includes: The forecast sequence of multiple meteorological factors within the preset time window is extracted from meteorological forecast data. The multiple meteorological factors include at least three of temperature, humidity, wind speed, air pressure, and irradiance. The extreme values, magnitude of change, and rate of change of each meteorological factor prediction sequence are calculated separately. Based on the calculation results, multiple single-factor extreme indices of multiple meteorological factors are evaluated. The multiple single-factor extreme indices are weighted and summed to obtain a comprehensive meteorological extreme index. When the comprehensive meteorological extreme index is greater than or equal to the preset extreme weather trigger threshold, the target area is determined to be an extreme weather scenario within the preset time window; When the comprehensive meteorological extreme index is less than the extreme weather trigger threshold, the target area is determined to be a non-extreme weather scenario within the preset time window.

8. The intelligent power load forecasting method based on meteorological assessment according to claim 1, characterized in that, The steps involved in constructing an extreme weather correction model include: Based on the power grid operation records of the target area, several sample single-factor extreme index sets and several sample power load prediction sequences are collected. The historical power load sequences of different sample single-factor extreme index sets and sample power load prediction sequences within the historical time window are collected as sample corrected power load prediction sequences, resulting in several sample corrected power load prediction sequences. The aforementioned sample single-factor extreme index sets, sample power load prediction sequences, and sample modified power load prediction sequences are used as training data and divided into P equal parts to obtain P modified training sets. The long short-term memory network is trained to convergence using the P corrected training sets, generating P extreme weather correction branches, which are then combined to obtain the extreme weather correction model.

9. A smart power load forecasting system based on meteorological assessment, characterized in that, The method for executing the intelligent power load forecasting method based on meteorological assessment as described in any one of claims 1-8 includes: The data acquisition and processing module is used to acquire multi-source prediction data of the target area within a preset time window, perform standardization and spatiotemporal alignment on the multi-source prediction data, and generate a multi-dimensional input tensor. The load forecasting module is used to assess the load forecasting complexity based on the multi-source forecasting data to determine the load forecasting complexity index, call the adaptive load forecaster based on the load forecasting complexity index, and obtain the power load forecasting result within a preset time window based on the multi-dimensional input tensor. The module calculates the meteorological mutation index based on the meteorological forecasting data and obtains the load forecasting complexity index based on the weighted assessment of the meteorological mutation index and the new energy output fluctuation index. The weather determination and correction module is used to determine whether the target area will experience non-extreme or extreme weather within a preset time window based on the multi-source forecast data. If it is non-extreme weather, the module outputs the power load forecast result; if it is extreme weather, the module calls the extreme weather correction model to correct the power load forecast result and outputs the corrected load forecast result, including: If it is extreme weather, the difference between the comprehensive meteorological extreme index and the extreme weather trigger threshold is calculated as the extreme index difference; Construct an extreme weather correction model, wherein the extreme weather correction model includes P extreme weather correction branches; The ratio of the extreme exponential difference to the preset benchmark extreme exponential deviation is multiplied by the initial number of branches and rounded down to obtain the number of adaptive branches L, where the initial number of branches is one-third P, and L is greater than or equal to 2 and less than or equal to P. L branches are randomly selected from the P extreme weather correction branches. Based on multiple single-factor extreme indices of multiple meteorological factors, the mean sequence of power load prediction in the power load prediction results is corrected, and L initial corrected power load prediction sequences are output. The mean is then calculated to obtain the corrected power load prediction sequence. For the L initial corrected power load prediction sequences, the corrected power load variation coefficient at the same time is calculated, and the average value of the L corrected power load variation coefficients at multiple times is calculated as the prediction correction fluctuation coefficient. The initial corrected prediction confidence level is obtained by multiplying the ratio of the predicted corrected volatility coefficient to the preset benchmark predicted corrected volatility coefficient by the preset benchmark corrected prediction confidence level and rounding down. The product of the current prediction confidence and the initial revised prediction confidence is used as the current revised prediction confidence. The corrected power load forecast sequence and the current corrected forecast confidence level are output as the corrected load forecast result.

Images