Method and system for generating medium and long term provincial load time series scenarios considering accumulated temperature effect under meteorological data driving

Abstract

The present application relates to the meteorological data driven under the consideration of the medium and long term provincial load time sequence scene generation method and system of accumulated temperature effect, belongs to the electric power system engineering technical field. For the meteorological sensitive season, the load decomposition theory is used to divide it into two parts of basic load and meteorological sensitive load, the peak value of basic load and meteorological sensitive load is estimated by using multivariate grey system model, and the proportion of meteorological sensitive load is estimated by combining the cumulative effect correction model of air temperature and the CNN-BiLSTM-Attention model of meteorological data, so as to obtain the load of meteorological sensitive season;At the same time, the load of non-meteorological sensitive season is generated. Finally, the meteorological sensitive load and the non-meteorological sensitive load are combined to obtain the provincial load scene. The present application discusses the load scene generation method by season, improves the correlation of load, economic population index and meteorological data, so that the generated scene has high reliability.

Description

Technical Field

[0001] This invention relates to a method and system for generating medium- and long-term provincial load time-series scenarios based on meteorological data and taking into account accumulated temperature effects, and belongs to the field of power system engineering technology. Background Technology

[0002] In recent years, with the integration of a high proportion of renewable energy into the power system, the impact of meteorology on load power has become increasingly severe, and the volatility and uncertainty of power load have become more pronounced. With the development of the national economy, the proportion of summer and winter loads has continued to rise, which has not only increased the load's sensitivity to temperature but also led to record-breaking summer and winter loads, posing a significant challenge to power supply security. Therefore, accurately generating future load time-series scenarios based on historical meteorological data, while considering economic development, is of great significance for power balance planning, peak-load power supply assurance, and power system planning.

[0003] Existing load scenario generation methods can be categorized into statistical methods and deep learning methods based on their modeling approaches. Statistical methods can be further divided into probabilistic model-based and time series-based methods. Probabilistic model-based methods involve statistically analyzing the probability distribution of historical load data and then sampling it to obtain the output scenario. However, probabilistic model-based methods require a large amount of sample data and rely on excessive simplifications and assumptions, potentially resulting in significant differences between the generated scenario and the actual scenario, failing to preserve the temporal order of the original data. Time series-based methods utilize the characteristics and patterns of time series data to generate scenarios, often employing Markov chains as a technique. They estimate the probability distribution of the next time step using a transition matrix, and use the expected value of this probability distribution as the load value at the current time step. However, they have high data quality requirements and struggle to generate comprehensive and accurate loads based on limited data. With the development of artificial intelligence, deep learning methods such as generative adversarial networks (GANs) and variational autoencoders (VAEs) introduce higher-dimensional feature information, such as meteorological data, user behavior patterns, or socioeconomic activity indicators. These methods can better capture the correlations between diverse data and describe the uncertainty of stochastic processes.

[0004] However, existing methods aim to generate load scenarios with similar patterns of change to historical load curves, but they suffer from insufficient correlation with factors such as socio-economic development and population consumption levels. With the increasing influence of socio-economic development and the growth of air conditioning load capacity, generating load scenarios solely through existing methods is insufficient to characterize future load changes and diversity.

[0005] Chinese patent document CN119067320A discloses a method and system for generating time-series load scenarios for provincial power grids under critical weather conditions for power supply assurance. While it generates these scenarios by considering economic development and employing load decomposition theory, it is limited to extreme weather events such as cold waves and heat waves. It does not address loads in spring and autumn and does not consider the cumulative effect of temperatures caused by consecutive days of high or low temperatures, thus failing to provide comprehensive and accurate load information for power system planning. Therefore, this invention is proposed. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for generating medium- and long-term provincial load time-series scenarios that take into account accumulated temperature effects under meteorological data.

[0007] The technical solution of the present invention is as follows:

[0008] A method for generating medium- to long-term provincial load time-series scenarios driven by meteorological data and taking into account accumulated temperature effects, the steps of which are as follows:

[0009] (1) Conduct a correlation analysis between load and meteorological factors, and divide the annual load sequence into meteorologically sensitive seasonal load (summer and winter) and non-meteorologically sensitive seasonal load (spring and autumn).

[0010] (2) For the load of the meteorologically sensitive season, the load is decomposed, and the temperature cumulative effect model, the multivariate grey system model and the CNN-BiLSTM-Attention model are constructed to generate the time series scene of the load of the meteorologically sensitive season.

[0011] (3) For loads in non-meteorologically sensitive seasons, select daily load characteristic indicators, perform load indicator probability distribution fitting and sampling, obtain load characteristic indicators for non-meteorologically sensitive seasons in the target year, and generate load time series scenarios for non-meteorologically sensitive seasons based on daily characteristic indicators.

[0012] (4) Combine the load of the meteorologically sensitive season with the load of the non-meteorologically sensitive season to obtain the target year load time sequence scenario.

[0013] According to a preferred embodiment of the present invention, in step (1), based on the 5-day moving average temperature sequence of the year, in the daily average temperature sequence (9 days) corresponding to the first five consecutive moving averages that meet a certain climate season index threshold, the first date greater than or equal to 10℃ is the start date of spring; the first date greater than or equal to 22℃ is the start date of summer; the first date less than 22℃ is the start date of autumn; the first date less than 10℃ is the start date of winter, and the day before the start date of each season is the end date of the previous season; the summer and winter loads are divided into meteorologically sensitive season loads, and the spring and autumn loads are divided into non-meteorologically sensitive season loads.

[0014] According to a preferred embodiment of the present invention, step (2) specifically includes the following steps:

[0015] (2-1) Select the average daily load of spring and autumn with a temperature of 16℃~25℃ and no rainfall in the historical years as the daily base load of spring and autumn in that year. Then take the average of the daily base load of spring and autumn in that year as the annual base load of that year. Subtract the provincial daily load curves of summer and winter from the annual base load curves one by one to obtain the meteorological sensitive load sequence of summer and winter. (The provincial daily load curves of summer and winter refer to the curves formed by the load power of the province at 24 points in summer and winter. The number of provincial daily load curves is determined by the number of days in summer and winter as defined in step (1). There is only one annual base load curve. Subtract the provincial daily load curves of summer and winter from the annual base load curves one by one to obtain the daily meteorological sensitive load sequence. Then splice them together day by day to obtain the meteorological sensitive load sequence of summer and winter.)

[0016] (2-2) Construct a multivariate grey system model of the relationship between peak meteorological sensitive load, annual base load and economic indicators;

[0017] (2-3) Construct a temperature cumulative effect correction model for daily maximum temperature and daily maximum meteorological sensitive load;

[0018] (2-4) Construct a CNN-BiLSTM-Attention model between the proportion of meteorological sensitive loads and meteorological data.

[0019] According to a preferred embodiment of the present invention, in step (2-1), the calculation formulas for the annual basic load and the meteorologically sensitive load sequence are as follows:

[0020] (1)

[0021] in, They are respectively the i-th day and j-th day of spring month a and autumn month c. Provincial load at any time Let be the annual base load value at time j, and m and n be the number of days with suitable temperatures in spring and autumn, respectively. These are the daily load curve values ​​for the province during summer and winter. This is a meteorologically sensitive load sequence.

[0022] The daily base load refers to the load values ​​at 24:00 throughout the day that meet the conditions of a temperature between 16℃ and 25℃ and no rainfall, forming the daily base load curve; the annual base load refers to the average of all the selected daily base loads to obtain the annual base load curve, which is formed by connecting the 24-hour annual base load values.

[0023] According to a preferred embodiment of the present invention, in step (2-2), the multivariate grey system model is:

[0024] (2)

[0025] In the formula, The original characteristic sequence is the historical annual basic load sequence or the peak sequence of meteorological sensitive loads. Original feature sequence A first-order AGO accumulator sequence; This is a sequence of influencing factors, corresponding to the historical socioeconomic indicators (including GDP, CPI, primary industry value, tertiary industry value, year-end total population, and residential electricity consumption), with subscripts... Indicates the first One economic and humanistic indicator, , The total number of indicators; for A first-order AGO accumulator sequence; These are the time values; a, b i c and d are model parameters, which are estimated using the least squares method; Background value coefficient;

[0026] Once the model training is complete, given the target year's economic indicators, the annual base load and meteorologically sensitive load peak values ​​for the target year can be generated.

[0027] According to a preferred embodiment of the present invention, in step (2-3), the temperature accumulation effect correction model is as follows:

[0028] (3)

[0029] In the formula: The highest temperature of the day to be predicted. , This is the limit temperature; Forecast date The highest temperature of the day; This represents the maximum cumulative number of days. This is the cumulative effect coefficient;

[0030] The temperature rise curve is used to determine the limit temperature. The temperature rise curve is a curve fitted to the daily maximum temperature and the daily maximum meteorological sensitive load (the daily maximum meteorological sensitive load is the maximum value of the daily meteorological sensitive load sequence). The temperature corresponding to the maximum slope of the curve is the limit temperature. When the limit temperature is exceeded, the temperature accumulation effect is considered. The maximum number of accumulation days is determined by trial and error. That is, the correlation coefficient between the meteorological sensitive load and the maximum temperature for different maximum accumulation days is calculated. The number of days corresponding to the maximum correlation coefficient is the maximum accumulation days.

[0031] The correlation coefficient is used to solve for k, satisfying the condition that the correlation coefficient between the corrected maximum temperature and the meteorological load is maximized. Under the selected maximum cumulative number of days p, the optimized solution model for the cumulative effect coefficient sequence is as follows:

[0032] (4)

[0033] (5)

[0034] in, The corrected highest temperature vector corresponds to the highest meteorological sensitive load vector. The constraint condition indicates that the corrected daily average temperature is calculated from the target date to the date before the target date. The weighted average of daily temperature data, and the closer to the target date, the stronger the impact of the daily temperature on the load;

[0035] Based on historical maximum temperature records and their corresponding 24-hour temperature sequences, a linear regression model is constructed between the daily maximum temperature of a province and the hourly temperatures within that province over 24 hours. The 24-hour temperatures of the province are then corrected point by point using the corrected daily maximum temperature. The linear regression model is as follows:

[0036] (6)

[0037] In the formula, It is the intercept term. It is the coefficient of the highest temperature. It is an error term. The highest temperature in the province on day i. For the first Heavenly The average temperature in the province over the hours , It can be estimated using the least squares method;

[0038] The correlation coefficient between the hourly temperature data of each city and the corresponding hourly provincial average temperature is calculated. Based on this, the hourly weighting coefficient of each city is calculated, and then the hourly temperature of each city is accurately corrected. The calculation formula is as follows:

[0039] (7)

[0040] (8)

[0041] (9)

[0042] In the formula, Indicates the first The city Hours and the province's first The correlation coefficient between hours Indicates the first Heavenly The city The temperature in the past hour Indicates the first The city The weighting coefficient for each hour Indicates the first Heavenly The city Temperature after 24 hours correction.

[0043] Equation (3) corrects the daily maximum temperature. Since the subsequent CNN-BiLSTM-Attention model is input to the temperature matrix composed of the hourly temperatures of each city, the corrected temperature matrix is ​​obtained by using equations (6)-(9) to correct the hourly temperatures of each city based on the daily maximum temperature corrected by equation (3).

[0044] According to a preferred embodiment of the present invention, in step (2-4), the CNN-BiLSTM-Attention model includes a convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism;

[0045] A convolutional neural network consists of convolutional layers, pooling layers, and fully connected layers. The function of the convolutional layer is feature extraction. The convolutional kernel slides on the feature map and performs convolution operations at each position to capture specific information contained between adjacent elements in the feature map, thereby effectively mining and extracting key feature information. The convolution operation formula is as follows:

[0046] (10)

[0047] In the formula, This represents the i-th convolutional kernel in convolutional layer l. Let J represent the j-th feature map input to convolutional layer l, where J represents the number of feature maps input to convolutional layer l. This represents the convolution operation. Here, f() represents the bias term, and f is the activation function of the convolutional layer l. This represents the i-th feature map output by convolutional layer l after the convolution operation;

[0048] After convolutional feature extraction, the output feature map is passed to the pooling layer for feature selection and information filtering. This process not only achieves information compression and dimensionality reduction, but also helps to prevent overfitting caused by excessive complexity of the network model.

[0049] The fully connected layer is located at the end of the convolutional neural network. It non-linearly combines the results calculated by the convolutional and pooling layers and transforms them into a one-dimensional vector output.

[0050] Bidirectional Long Short-Term Memory (BiLSTM) network consists of a forward LSTM and a backward LSTM. The forward LSTM processes the input sequence in the forward direction, while the backward LSTM processes the sequence in the reverse direction. After processing, the outputs of the two LSTMs are concatenated to obtain the output of the Bidirectional Long Short-Term Memory network. When processing time series data, it can effectively capture the dependencies between variables. For the input at time t, the calculation formula is as follows:

[0051] (11)

[0052] (12)

[0053] (13)

[0054] In the formula, , These are the outputs for forward and backward propagation at the current time, respectively; , These are the outputs of the forward and backward propagation from the previous time step, respectively; , , , , , As weight; For bias;

[0055] An attention mechanism layer is added after the bidirectional long short-term memory network. The attention mechanism layer assigns different weights to each feature and calculates the weights of the attention scores through the softmax function. Finally, the sequence of meteorological sensitive load proportions is output.

[0056] Integrating an attention mechanism into the Bi-LSTM network architecture aims to focus on the information features processed by the Bi-LSTM layers, and to accurately extract information that is crucial to the prediction task by dynamically allocating weight ratios.

[0057] The formula for calculating the attention mechanism is as follows:

[0058] (14)

[0059] In the formula, This is the feature matrix output by the bidirectional long short-term memory network; Characteristic matrix through The result of the function operation; This is the weight matrix; This is the normalized weight matrix; The output vector after weights have been assigned by the attention mechanism layer;

[0060] Finally, the hourly weather data of each city and the corresponding meteorological sensitive load ratio (the meteorological sensitive load ratio refers to the meteorological sensitive load value divided by the peak meteorological sensitive load of that year) are input into the CNN-BiLSTM-Attention model. The model is trained using a typical year (the typical year is the year closest to the target year, and it is assumed that the load characteristics of the typical year are closest to the target year, so the typical year is used for training) to obtain the mapping relationship between meteorological data and meteorological sensitive load ratio. Based on this mapping relationship, the meteorological sensitive load ratio of the target year is generated using historical weather data. The meteorological sensitive load ratio is multiplied by the peak meteorological sensitive load, and then the annual basic load is added to obtain the meteorological sensitive seasonal load of the target year.

[0061] According to a preferred embodiment of the present invention, step (3) specifically includes the following steps:

[0062] (3-1) Statistical analysis of historical load characteristics during non-meteorologically sensitive seasons, including daily average load, peak load, peak-to-valley difference, peak load rate, and valley load rate;

[0063] (3-2) Construct the joint probability density function of daily load characteristic indicators using the Copula function and perform random sampling;

[0064] (3-3) Analyze load characteristic indicators and estimate the target year for load characteristic indicators that change regularly;

[0065] (3-4) Based on the daily load characteristic index obtained by the above method, select the daily load sequence with the smallest Euclidean distance between the historical load sequence and the generated daily load characteristic index as the benchmark load sequence, construct an optimization model, optimize and adjust the benchmark load sequence, and obtain the target load curve.

[0066] According to a preferred embodiment of the present invention, in step (3-2), the Copula function can accurately describe the correlation between multivariate random variables. Its main idea is to decompose the joint distribution function of the multivariate variables into multiple marginal distribution functions, and then connect them through a suitable Copula function. As can be seen from Sklar's theorem: if the random variables... The marginal cumulative distribution function (CDF) is If the joint distribution function is G, then there must exist a copula function C such that:

[0067] (15)

[0068] In the above formula, For each daily load characteristic index, there are random variables, specifically the daily average load, peak load, peak-to-valley difference, peak load rate, and valley load rate. The corresponding marginal cumulative distribution function; the joint probability density function is the combination of the various marginal cumulative distribution functions using the Copula function, as shown in the above formula. Random sampling refers to randomly selecting multiple sets of daily load characteristic indicators. Each set of daily load characteristic indicators represents the daily load characteristic indicators of a certain day. The extracted daily load characteristic indicators are then used to optimize the base load for subsequent years.

[0069] According to a preferred embodiment of the present invention, in step (3-3), by statistically analyzing the historical load characteristic index changes, it is found that the daily average load, daily peak load, and daily peak-to-valley load difference show a significant upward trend over time, while the peak load rate and valley load rate show no obvious regular changes. Therefore, the median changes of these three load characteristic indices for the target year are generated using a multivariate grey system model, and a normal distribution with median deviation as the mean and standard deviation of 1 is constructed to obtain the load characteristic index for the target year. The distribution of load characteristic index X is expressed as follows:

[0070] (16)

[0071] in, The median deviation is the difference between the median of the target year's load characteristic index and the median of the typical year. Let the variance be, and .

[0072] According to a preferred embodiment of the present invention, in step (3-4), the optimization model is as follows:

[0073] (17)

[0074] In the formula: L(t) is the generated daily load sequence; The selected baseline load sequence;

[0075] The basic idea behind optimizing model constraints is that the daily characteristic indices of the generated sequence are consistent with the daily characteristic indices obtained from sampling, that is:

[0076]

[0077]

[0078]

[0079]

[0080]

[0081] In the formula, These correspond to the daily average load, daily peak load, daily peak-to-valley load difference, daily peak load rate, and daily valley load rate, respectively. and These represent the times when the maximum and minimum loads occur in the baseline load curve, respectively. and These represent the peak load period and the number of peak load periods, respectively. and These represent the periods of load trough and the number of periods of load trough, respectively. The above optimization model is a quadratic programming model. Solving this optimization problem will yield the optimal daily load sequence that meets the load index requirements.

[0082] The peak load periods are defined as 10:00 AM to 12:00 PM and 5:00 PM to 9:00 PM, and the number of peak load periods is the number of hours contained in the peak load periods, which is 8. The off-peak load periods are defined as 1:00 AM to 5:00 AM and 11:00 PM to 12:00 AM, and the number of off-peak load periods is the number of hours contained in the off-peak load periods, which is 7.

[0083] A system for generating medium- to long-term provincial load time-series scenarios based on meteorological data and taking into account accumulated temperature effects, including:

[0084] The analysis module is used to perform correlation analysis between load and meteorological factors, and to divide the annual load sequence into meteorologically sensitive season load and non-meteorologically sensitive season load;

[0085] The module for generating time-series load scenarios for meteorologically sensitive seasons is used to decompose the load for meteorologically sensitive seasons, construct a temperature cumulative effect model, a multivariate grey system model, and a CNN-BiLSTM-Attention model, and generate time-series load scenarios for meteorologically sensitive seasons.

[0086] The non-meteorologically sensitive season load time series scenario generation module is used to select daily load characteristic indicators for non-meteorologically sensitive season loads, perform load indicator probability distribution fitting and sampling, obtain non-meteorologically sensitive season load characteristic indicators for the target year, and generate non-meteorologically sensitive season load time series scenarios based on daily characteristic indicators.

[0087] The target year load time series scenario generation module is used to combine the load of meteorologically sensitive seasons with the load of non-meteorologically sensitive seasons to obtain the target year load time series scenario.

[0088] The beneficial effects of this invention are as follows:

[0089] This invention fully considers the impact of meteorological factors on load and classifies and models load characteristics in conjunction with the correlation between meteorological factors. By comprehensively using multiple models, load can be predicted more accurately, and the generated medium- and long-term load time-series scenarios are more consistent with reality, providing a more reliable basis for medium- and long-term planning in the power system and improving the stability and economy of power system operation. Attached Figure Description

[0090] Figure 1 This is a flowchart illustrating Embodiment 1 of the present invention;

[0091] Figure 2 This is a flowchart of the multivariable grey system model of the present invention;

[0092] Figure 3 This is a schematic diagram of the CNN-BiLSTM-Attention model of the present invention;

[0093] Figure 4 This is a per-unit curve of the average daily maximum load in Embodiment 1 of the present invention;

[0094] Figure 5 This is a comparison chart of the Pearson correlation coefficients between the daily maximum load and meteorological factors in Embodiment 1 of the present invention. Figure 5 (a) shows a comparison of correlation coefficients in spring. Figure 5 (b) is a comparison chart of correlation coefficients in summer. Figure 5 (c) is a comparison chart of correlation coefficients in autumn. Figure 5 The middle (d) chart shows the correlation coefficient comparison for winter.

[0095] Figure 6 This is a time-series diagram of the basic load of Embodiment 1 of the present invention;

[0096] Figure 7 This is a graph showing the peak value estimation results of meteorological sensitive loads in Embodiment 1 of the present invention;

[0097] Figure 8 This is a time-series load result diagram for a certain day during the meteorologically sensitive season (summer) of this invention;

[0098] Figure 9 This is a time-series load result diagram for a certain day during the meteorologically sensitive season (winter) of this invention;

[0099] Figure 10 This is a PDF comparison chart of the annual load for the target year in Embodiment 1 of the present invention;

[0100] Figure 11 This is a comparison chart of the annual load CDF for the target year in Embodiment 1 of the present invention. Detailed Implementation

[0101] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0102] Example 1:

[0103] like Figure 1 As shown in the figure, this embodiment provides a method for generating medium- and long-term provincial load time series scenarios that take into account accumulated temperature effects under meteorological data. The steps are as follows:

[0104] (1) Conduct a correlation analysis between load and meteorological factors, and divide the annual load sequence into meteorologically sensitive seasonal load (summer and winter) and non-meteorologically sensitive seasonal load (spring and autumn). Based on the 5-day moving average temperature sequence of the year, in the daily average temperature sequence (9 days) corresponding to the first 5 consecutive moving averages that meet a certain climate seasonal index threshold, the first date greater than or equal to 10℃ is the start date of spring; the first date greater than or equal to 22℃ is the start date of summer; the first date less than 22℃ is the start date of autumn; the first date less than 10℃ is the start date of winter, and the day before the start date of each season is the end date of the previous season. The summer and winter loads are divided into meteorologically sensitive seasonal loads, and the spring and autumn loads are divided into non-meteorologically sensitive seasonal loads.

[0105] (2) For the load in the meteorologically sensitive season, the load is decomposed, and a temperature cumulative effect model, a multivariate grey system model and a CNN-BiLSTM-Attention model are constructed to generate the time series scene of the load in the meteorologically sensitive season. The specific steps are as follows:

[0106] (2-1) Select the average daily load of spring and autumn with a temperature of 16℃~25℃ and no rainfall in the historical years as the daily base load of spring and autumn in that year. Then take the average of the daily base load of spring and autumn in that year as the annual base load of that year. Subtract the provincial daily load curves of summer and winter from the annual base load curves one by one to obtain the meteorological sensitive load sequence of summer and winter. (The provincial daily load curves of summer and winter refer to the curves formed by the load power of the province at 24 points in summer and winter. The number of provincial daily load curves is determined by the number of days in summer and winter as defined in step (1). There is only one annual base load curve. Subtract the provincial daily load curves of summer and winter from the annual base load curves one by one to obtain the daily meteorological sensitive load sequence. Then splice them together day by day to obtain the meteorological sensitive load sequence of summer and winter.)

[0107] The formulas for calculating the annual basic load and meteorological sensitive load sequences are as follows:

[0108] (1)

[0109] in, They are respectively the i-th day and j-th day of spring month a and autumn month c. Provincial load at any time Let be the annual base load value at time j, and m and n be the number of days with suitable temperatures in spring and autumn, respectively. These are the daily load curve values ​​for the province during summer and winter. This is a meteorologically sensitive load sequence.

[0110] The daily base load refers to the load values ​​at 24:00 throughout the day that meet the conditions of a temperature between 16℃ and 25℃ and no rainfall, forming the daily base load curve; the annual base load refers to the average of all the selected daily base loads to obtain the annual base load curve, which is formed by connecting the 24-hour annual base load values.

[0111] (2-2) Construct a multivariate grey system model of the relationship between meteorological sensitive load peak, annual base load and economic indicators, such as Figure 2 As shown;

[0112] The multivariate grey system model is as follows:

[0113] (2)

[0114] In the formula, The original characteristic sequence is the historical annual basic load sequence or the peak sequence of meteorological sensitive loads. Original feature sequence A first-order AGO accumulator sequence; This is a sequence of influencing factors, corresponding to the historical socioeconomic indicators (including GDP, CPI, primary industry value, tertiary industry value, year-end total population, and residential electricity consumption), with subscripts... Indicates the first One economic and humanistic indicator, , The total number of indicators; for A first-order AGO accumulator sequence; The time value; These are the background value coefficients; a, b i c and d are model parameters. For the parameter matrix, The least squares method is used for estimation, which satisfies... ,in:

[0115]

[0116]

[0117] The sequence is then obtained by cumulative subtraction using the time response formula. The simulated values, based on The result determines that the original sequence The background value coefficient with the smallest relative simulation error mean Then, the multivariate grey prediction model at this time is said to have the optimal background value coefficient. The OBGM(1,N) model.

[0118] Once the model training is complete, given historical annual national economic indicators, historical annual base load, and historical annual meteorological sensitive load peak values, the multivariate grey system is used to estimate the base load and meteorological sensitive load peak values ​​for future target years.

[0119] (2-3) Construct a temperature cumulative effect correction model for daily maximum temperature and daily maximum meteorological sensitive load;

[0120] The temperature accumulation effect correction model is as follows:

[0121] (3)

[0122] In the formula: The highest temperature of the day to be predicted. , This is the limit temperature; Forecast date The highest temperature of the day; This represents the maximum cumulative number of days. This is the cumulative effect coefficient;

[0123] The temperature rise curve is used to determine the threshold temperature, which is the temperature point where the load is most sensitive to temperature changes. A temperature rise curve is obtained by fitting the daily maximum temperature to the daily maximum temperature-sensitive load. The threshold temperature is the temperature corresponding to the maximum rate of change of the temperature rise curve. The calculation formula is as follows:

[0124]

[0125] In the formula, For temperature-sensitive loads, For temperature, The rate of change of the temperature rise curve;

[0126] Since the intensity of the cumulative effect varies depending on the duration of high temperatures, a trial-and-error method was used, starting from 1 and gradually increasing the maximum cumulative number of days. The correlation between the peak load and the peak temperature was calculated considering different cumulative numbers of days, and the maximum cumulative number of days that maximized the correlation was determined. .

[0127] Under the selected maximum cumulative number of days p, the optimization solution model for the cumulative effect coefficient sequence is as follows:

[0128] (4)

[0129] (5)

[0130] in, The corrected highest temperature vector corresponds to the highest meteorological sensitive load vector. The constraint condition indicates that the corrected daily average temperature is calculated from the target date to the date before the target date. The weighted average of daily temperature data, and the closer to the target date, the stronger the impact of the daily temperature on the load;

[0131] Based on historical maximum temperature records and their corresponding 24-hour temperature sequences, a linear regression model is constructed between the daily maximum temperature of a province and the hourly temperatures within that province over 24 hours. The 24-hour temperatures of the province are then corrected point by point using the corrected daily maximum temperature. The linear regression model is as follows:

[0132] (6)

[0133] In the formula, It is the intercept term. It is the coefficient of the highest temperature. It is an error term. The highest temperature in the province on day i. For the first Heavenly The average temperature in the province over the hours , It can be estimated using the least squares method;

[0134] The correlation coefficient between the hourly temperature data of each city and the corresponding hourly provincial average temperature is calculated. Based on this, the hourly weighting coefficient of each city is calculated, and then the hourly temperature of each city is accurately corrected. The calculation formula is as follows:

[0135] (7)

[0136] (8)

[0137] (9)

[0138] In the formula, Indicates the first The city Hours and the province's first The correlation coefficient between hours Indicates the first Heavenly The city The temperature in the past hour Indicates the first The city The weighting coefficient for each hour Indicates the first Heavenly The city Temperature after 24 hours correction.

[0139] Equation (3) corrects the daily maximum temperature. Since the subsequent CNN-BiLSTM-Attention model is input to the temperature matrix composed of the hourly temperatures of each city, the corrected temperature matrix is ​​obtained by using equations (6)-(9) to correct the hourly temperatures of each city based on the daily maximum temperature corrected by equation (3).

[0140] (2-4) Construct a CNN-BiLSTM-Attention model between the proportion of meteorological sensitive loads and meteorological data, such as... Figure 3 As shown.

[0141] The CNN-BiLSTM-Attention model includes a convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism;

[0142] A convolutional neural network consists of convolutional layers, pooling layers, and fully connected layers. The function of the convolutional layer is feature extraction. The convolutional kernel slides on the feature map and performs convolution operations at each position to capture specific information contained between adjacent elements in the feature map, thereby effectively mining and extracting key feature information. The convolution operation formula is as follows:

[0143] (10)

[0144] In the formula, This represents the i-th convolutional kernel in convolutional layer l. Let J represent the j-th feature map input to convolutional layer l, where J represents the number of feature maps input to convolutional layer l. This represents the convolution operation. Here, f() represents the bias term, and f is the activation function of the convolutional layer l. This represents the i-th feature map output by convolutional layer l after the convolution operation;

[0145] After convolutional feature extraction, the output feature map is passed to the pooling layer for feature selection and information filtering. This process not only achieves information compression and dimensionality reduction, but also helps to prevent overfitting caused by excessive complexity of the network model.

[0146] The fully connected layer is located at the end of the convolutional neural network. It non-linearly combines the results calculated by the convolutional and pooling layers and transforms them into a one-dimensional vector output.

[0147] Bidirectional Long Short-Term Memory (BiLSTM) network consists of a forward LSTM and a backward LSTM. The forward LSTM processes the input sequence in the forward direction, while the backward LSTM processes the sequence in the reverse direction. After processing, the outputs of the two LSTMs are concatenated to obtain the output of the Bidirectional Long Short-Term Memory network. When processing time series data, it can effectively capture the dependencies between variables. For the input at time t, the calculation formula is as follows:

[0148] (11)

[0149] (12)

[0150] (13)

[0151] In the formula, , These are the outputs for forward and backward propagation at the current time, respectively; , These are the outputs of the forward and backward propagation from the previous time step, respectively; , , , , , As weight; For bias;

[0152] An attention mechanism layer is added after the bidirectional long short-term memory network. The attention mechanism layer assigns different weights to each feature and calculates the weights of the attention scores through the softmax function. Finally, the sequence of meteorological sensitive load proportions is output.

[0153] Integrating an attention mechanism into the Bi-LSTM network architecture aims to focus on the information features processed by the Bi-LSTM layers, and to accurately extract information that is crucial to the prediction task by dynamically allocating weight ratios.

[0154] The formula for calculating the attention mechanism is as follows:

[0155] (14)

[0156] In the formula, This is the feature matrix output by the bidirectional long short-term memory network; Characteristic matrix through The result of the function operation; This is the weight matrix; This is the normalized weight matrix; The output vector after weights have been assigned by the attention mechanism layer;

[0157] Finally, the hourly weather data of each city and the corresponding meteorological sensitive load ratio (the meteorological sensitive load ratio refers to the meteorological sensitive load value divided by the peak meteorological sensitive load of that year) are input into the CNN-BiLSTM-Attention model. The model is trained using a typical year (the typical year is the year closest to the target year, and it is assumed that the load characteristics of the typical year are closest to the target year, so the typical year is used for training) to obtain the mapping relationship between meteorological data and meteorological sensitive load ratio. Based on this mapping relationship, the meteorological sensitive load ratio of the target year is generated using historical weather data. The meteorological sensitive load ratio is multiplied by the peak meteorological sensitive load, and then the annual basic load is added to obtain the meteorological sensitive seasonal load of the target year.

[0158] (3) For loads in non-meteorologically sensitive seasons, select daily load characteristic indicators, perform probability distribution fitting and sampling of load indicators, obtain load characteristic indicators for non-meteorologically sensitive seasons in the target year, and generate time series scenarios of loads in non-meteorologically sensitive seasons based on daily characteristic indicators. The specific steps are as follows:

[0159] (3-1) Statistical analysis of historical load characteristics during non-meteorologically sensitive seasons, including daily average load, peak load, peak-to-valley difference, peak load rate, and valley load rate;

[0160] (3-2) Construct the joint probability density function of daily load characteristic indicators using the Copula function and perform random sampling;

[0161] Copula functions can accurately describe the correlation between multivariate random variables. The main idea is to decompose the joint distribution function of the multivariate variables into multiple marginal distribution functions, and then connect them using a suitable Copula function. According to Sklar's theorem, if the random variables... The marginal cumulative distribution function (CDF) is If the joint distribution function is G, then there must exist a copula function C such that:

[0162] (15)

[0163] In the above formula, For each daily load characteristic index, there are random variables, specifically the daily average load, peak load, peak-to-valley difference, peak load rate, and valley load rate. The corresponding marginal cumulative distribution function; the joint probability density function is the combination of the various marginal cumulative distribution functions using the Copula function, as shown in the above formula. Random sampling refers to randomly selecting multiple sets of daily load characteristic indicators. Each set of daily load characteristic indicators represents the daily load characteristic indicators of a certain day. The extracted daily load characteristic indicators are used to optimize the base load for subsequent years.

[0164] (3-3) Analyze load characteristic indicators and estimate the target year for load characteristic indicators that change regularly;

[0165] By statistically analyzing the historical load characteristic indicators, it was found that the daily average load, daily peak load, and daily peak-to-valley load difference show a significant upward trend over time, while the peak load rate and valley load rate show no obvious regularity. Therefore, a multivariate grey system model was used to generate the median changes of these three load characteristic indicators for the target year. A normal distribution with median deviation as the mean and standard deviation of 1 was constructed, thus obtaining the load characteristic indicators for the target year. The distribution of load characteristic indicator X is expressed as follows:

[0166] (16)

[0167] in, The median deviation is the difference between the median of the target year's load characteristic index and the median of the typical year. Let the variance be, and .

[0168] (3-4) Based on the daily load characteristic index obtained by the above method, select the daily load sequence with the smallest Euclidean distance between the historical load sequence and the generated daily load characteristic index as the benchmark load sequence, construct an optimization model, optimize and adjust the benchmark load sequence, and obtain the target load curve.

[0169] The optimized model is as follows:

[0170] (17)

[0171] In the formula: L(t) is the generated daily load sequence; The selected baseline load sequence;

[0172] The basic idea behind optimizing model constraints is that the daily characteristic indices of the generated sequence are consistent with the daily characteristic indices obtained from sampling, that is:

[0173]

[0174]

[0175]

[0176]

[0177]

[0178] In the formula, These correspond to the daily average load, daily peak load, daily peak-to-valley load difference, daily peak load rate, and daily valley load rate, respectively. and These represent the times when the maximum and minimum loads occur in the baseline load curve, respectively. and These represent the peak load period and the number of peak load periods, respectively. and These represent the periods of load trough and the number of periods of load trough, respectively. The above optimization model is a quadratic programming model. Solving this optimization problem will yield the optimal daily load sequence that meets the load index requirements.

[0179] (4) Combine the load of the meteorologically sensitive season with the load of the non-meteorologically sensitive season to obtain the target year load time sequence scenario.

[0180] To verify the technical concept of this embodiment, a specific calculation example analysis is given below.

[0181] Taking a certain province as the research object, we collected historical socio-economic indicators (including GDP, CPI, value of primary industry, value of secondary industry, value of tertiary industry, residential electricity consumption and year-end total population) from 2017 to 2022, provincial load data, and meteorological element data of all cities under its jurisdiction (hourly load and meteorological data of 16 cities).

[0182] Based on the calculation example analysis, the per-unit curve of the average daily maximum load from 2017 to 2022 is as follows: Figure 4 As shown, the load fluctuations are stronger and the peak values ​​are higher during meteorologically sensitive seasons, while the load fluctuations are weaker during non-meteorologically sensitive seasons. Therefore, it is necessary to divide the load into seasons. The Pearson correlation coefficients between the daily maximum load and meteorological elements under both traditional and climatic seasonal divisions are calculated, and the results are as follows: Figure 5 As shown in Table 1, the Pearson correlation coefficients between the base load, peak meteorological sensitive load, and economic and demographic indicators were calculated. Factors with strong correlations were selected as the correlation sequences for the model.

[0183] Table 1: Pearson correlation coefficients between basic load, peak values ​​of meteorological sensitive loads, and socioeconomic development indicators.

[0184]

[0185] As shown in Table 1, the multivariate grey system model in the example will select annual GDP, CPI, primary industry, tertiary industry, residential electricity consumption, and year-end total population as relevant series. The basic load time series results obtained from this model are as follows: Figure 6 As shown; peak values ​​of meteorological sensitive loads are as follows: Figure 7 As shown.

[0186] Multiplying the peak value of the meteorological sensitive load by the proportion of the meteorological sensitive load and then adding the base load yields the load time series result for a specific day in the meteorological sensitive season (summer), as shown below. Figure 8 As shown; the load time series results for a certain day during the meteorologically sensitive season (winter) are as follows. Figure 9 As shown.

[0187] The comparison results of the full-year load PDF in 2022 are as follows: Figure 10 As shown; the target year's annual load CDF comparison results are as follows. Figure 11 As shown in the figure, the results demonstrate that, driven by socioeconomic indicators and meteorological data, the method in this embodiment can accurately reflect changes in load.

[0188] Example 2:

[0189] This embodiment provides a system for generating medium- and long-term provincial load time-series scenarios based on meteorological data and considering accumulated temperature effects, including:

[0190] The analysis module is used to perform correlation analysis between load and meteorological factors, and to divide the annual load sequence into meteorologically sensitive season load and non-meteorologically sensitive season load;

[0191] The module for generating time-series load scenarios for meteorologically sensitive seasons is used to decompose the load for meteorologically sensitive seasons, construct a temperature cumulative effect model, a multivariate grey system model, and a CNN-BiLSTM-Attention model, and generate time-series load scenarios for meteorologically sensitive seasons.

[0192] The non-meteorologically sensitive season load time series scenario generation module is used to select daily load characteristic indicators for non-meteorologically sensitive season loads, perform load indicator probability distribution fitting and sampling, obtain non-meteorologically sensitive season load characteristic indicators for the target year, and generate non-meteorologically sensitive season load time series scenarios based on daily characteristic indicators.

[0193] The target year load time series scenario generation module is used to combine the load of meteorologically sensitive seasons with the load of non-meteorologically sensitive seasons to obtain the target year load time series scenario.

[0194] Those skilled in the art will recognize that the units, i.e., algorithm steps, of the various examples described in connection with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0195] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for generating medium- to long-term provincial load time-series scenarios based on meteorological data and considering accumulated temperature effects, characterized in that... The steps are as follows: (1) Conduct a correlation analysis between load and meteorological factors, and divide the annual load sequence into meteorologically sensitive season load and non-meteorologically sensitive season load; (2) For the load in the meteorologically sensitive season, the load is decomposed, and a temperature cumulative effect model, a multivariate grey system model and a CNN-BiLSTM-Attention model are constructed to generate the time series scene of the load in the meteorologically sensitive season. The specific steps are as follows: (2-1) Select the average daily load of spring and autumn when the temperature is between 16℃ and 25℃ and there is no rainfall in the historical years as the daily base load of spring and autumn in that year. Then take the average of the daily base load of spring and autumn in that year as the annual base load of that year. Subtract the provincial daily load curve and the annual base load curve one by one to obtain the summer and winter meteorological sensitive load sequence. (2-2) Construct a multivariate grey system model of the relationship between peak meteorological sensitive load, annual base load and economic indicators; (2-3) Construct a temperature cumulative effect correction model for daily maximum temperature and daily maximum meteorological sensitive load; The temperature accumulation effect correction model is as follows: （3） In the formula: The highest temperature of the day to be predicted. , This is the limit temperature; Forecast date The highest temperature of the day; This represents the maximum cumulative number of days. This is the cumulative effect coefficient; The temperature rise curve is used to determine the limit temperature. The temperature rise curve is a curve fitted by the daily maximum temperature and the daily maximum meteorological sensitive load. The temperature corresponding to the maximum slope of the curve is the limit temperature. When the limit temperature is exceeded, the temperature accumulation effect is considered. The maximum number of accumulation days is determined by trial and error. That is, the correlation coefficient between the meteorological sensitive load and the maximum temperature for different maximum accumulation days is calculated. The number of days corresponding to the maximum correlation coefficient is the maximum accumulation days. The correlation coefficient is used to solve for k, satisfying the condition that the correlation coefficient between the corrected maximum temperature and the meteorological load is maximized. Under the selected maximum cumulative number of days p, the optimized solution model for the cumulative effect coefficient sequence is as follows: （4） （5） in, The corrected highest temperature vector corresponds to the highest meteorological sensitive load vector. The constraint condition indicates that the corrected daily average temperature is calculated from the target date to the date before the target date. The weighted average of daily temperature data, and the closer to the target date, the stronger the impact of the daily temperature on the load; Based on historical maximum temperature records and their corresponding 24-hour temperature sequences, a linear regression model is constructed between the daily maximum temperature of a province and the hourly temperatures within that province over 24 hours. The 24-hour temperatures of the province are then corrected point by point using the corrected daily maximum temperature. The linear regression model is as follows: （6） In the formula, It is the intercept term. It is the coefficient of the highest temperature. It is an error term. The highest temperature in the province on day i. For the first Heavenly The average temperature in the province over the hours; The correlation coefficient between the hourly temperature data of each city and the corresponding hourly provincial average temperature is calculated. Based on this, the hourly weighting coefficient of each city is calculated, and then the hourly temperature of each city is accurately corrected. The calculation formula is as follows: （7） （8） （9） In the formula, Indicates the first The city Hours and the province's first The correlation coefficient between hours Indicates the first Heavenly The city The temperature in the past hour Indicates the first The city The weighting coefficient for each hour Indicates the first Heavenly The city Temperature corrected hourly; (2-4) Construct a CNN-BiLSTM-Attention model between the proportion of meteorological sensitive loads and meteorological data; (3) For loads in non-meteorologically sensitive seasons, select daily load characteristic indicators, perform load indicator probability distribution fitting and sampling, obtain load characteristic indicators for non-meteorologically sensitive seasons in the target year, and generate load time series scenarios for non-meteorologically sensitive seasons based on daily characteristic indicators. (4) Combine the load of the meteorologically sensitive season with the load of the non-meteorologically sensitive season to obtain the target year load time sequence scenario.

2. The method for generating medium- and long-term provincial load time-series scenarios driven by meteorological data and taking into account accumulated temperature effects as described in claim 1, characterized in that, In step (1), based on the 5-day moving average temperature sequence of the year, in the daily average temperature sequence corresponding to the first 5 consecutive moving averages that meet a certain climate season index threshold, the first date greater than or equal to 10℃ is the start date of spring; the first date greater than or equal to 22℃ is the start date of summer; the first date less than 22℃ is the start date of autumn; the first date less than 10℃ is the start date of winter, and the day before the start date of each season is the end date of the previous season; the summer and winter loads are divided into meteorologically sensitive season loads, and the spring and autumn loads are divided into non-meteorologically sensitive season loads.

3. The method for generating medium- and long-term provincial load time-series scenarios driven by meteorological data and taking into account accumulated temperature effects as described in claim 1, characterized in that, In step (2-1), the calculation formulas for the annual basic load and the meteorological sensitive load sequence are as follows: （1） in, The load data for each province are defined as follows: (a) in spring and (c) in autumn, on day i at time j, where j = 1, 2, ..., 24. Let be the annual base load value at time j, and m and n be the number of days with suitable temperatures in spring and autumn, respectively. These are the daily load curve values ​​for the province during summer and winter. This is a meteorologically sensitive load sequence.

4. The method for generating medium- and long-term provincial load time-series scenarios driven by meteorological data and taking into account accumulated temperature effects as described in claim 3, characterized in that, In step (2-2), the multivariate grey system model is as follows: （2） In the formula, The original characteristic sequence is the historical annual basic load sequence or the peak sequence of meteorological sensitive loads. Original feature sequence A first-order AGO accumulator sequence; This is the sequence of influencing factors, that is, the sequence of socio-economic indicators corresponding to historical years. The subscript m represents the m-th economic and humanistic indicator, m=1,2,...,N, where N is the total number of indicators. for A first-order AGO accumulator sequence; These are the time values; a, b i c and d are model parameters, which are estimated using the least squares method; Background value coefficient; Once the model training is complete, given the target year's economic indicators, the annual base load and meteorologically sensitive load peak values ​​for the target year can be generated.

5. The method for generating medium- and long-term provincial load time-series scenarios driven by meteorological data and taking into account accumulated temperature effects as described in claim 4, characterized in that, In steps (2-4), the CNN-BiLSTM-Attention model includes a convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism; A convolutional neural network consists of convolutional layers, pooling layers, and fully connected layers. The function of the convolutional layer is feature extraction. The convolutional kernel slides on the feature map and performs convolution operations at each position to capture specific information contained between adjacent elements in the feature map, thereby effectively mining and extracting key feature information. The convolution operation formula is as follows: （10） In the formula, This represents the p-th convolutional kernel in convolutional layer l. Let J represent the j-th feature map input to convolutional layer l, where J represents the number of feature maps input to convolutional layer l. This represents the convolution operation. Here, f() represents the bias term, and f is the activation function of the convolutional layer l. This represents the p-th feature map output by convolutional layer l after the convolution operation; After convolutional feature extraction, the output feature map is passed to the pooling layer for feature selection and information filtering; The fully connected layer is located at the end of the convolutional neural network. It non-linearly combines the results calculated by the convolutional and pooling layers and transforms them into a one-dimensional vector output. The bidirectional long short-term memory network consists of a forward LSTM and a backward LSTM. The forward LSTM processes the input sequence in the forward direction, while the backward LSTM processes the sequence in the reverse direction. After processing, the outputs of the two LSTMs are concatenated to obtain the output of the bidirectional long short-term memory network. When processing time series data, it captures the dependencies between variables. For the input at time t, the calculation formula is as follows: （11） （12） （13） In the formula, , These are the outputs for forward and backward propagation at the current time, respectively; , These are the outputs of the forward and backward propagation from the previous time step, respectively; , , , , , As weight; For bias; An attention mechanism layer is added after the bidirectional long short-term memory network. The attention mechanism layer assigns different weights to each feature and calculates the weights of the attention scores through the softmax function. Finally, the sequence of meteorological sensitive load proportions is output. The formula for calculating the attention mechanism is as follows: （14） In the formula, This is the feature matrix output by the bidirectional long short-term memory network; Characteristic matrix through The result of the function operation; This is the weight matrix; This is the normalized weight matrix; The output vector after weights have been assigned by the attention mechanism layer; Finally, the hourly weather data of each city and the corresponding meteorological sensitive load percentage are input into the CNN-BiLSTM-Attention model. The model is trained using a typical year to obtain the mapping relationship between meteorological data and meteorological sensitive load percentage. Based on this mapping relationship, the meteorological sensitive load percentage of the target year is generated using historical weather data. The meteorological sensitive load percentage is multiplied by the peak value of the meteorological sensitive load, and then the annual basic load is added to obtain the meteorological sensitive seasonal load of the target year.

6. The method for generating medium- and long-term provincial load time-series scenarios driven by meteorological data and taking into account accumulated temperature effects as described in claim 5, characterized in that, In step (3), the specific steps are as follows: (3-1) Statistical analysis of historical load characteristics during non-meteorologically sensitive seasons, including daily average load, peak load, peak-to-valley difference, peak load rate, and valley load rate; (3-2) Construct the joint probability density function of daily load characteristic indicators using the Copula function and perform random sampling; (3-3) Analyze load characteristic indicators and estimate the target year for load characteristic indicators that change regularly; (3-4) Based on the obtained daily load characteristic index, select the daily load sequence with the smallest Euclidean distance from the historical load sequence as the benchmark load sequence, construct an optimization model, optimize and adjust the benchmark load sequence, and obtain the target load curve.

7. The method for generating medium- and long-term provincial load time-series scenarios driven by meteorological data and taking into account accumulated temperature effects as described in claim 6, characterized in that, In step (3-2), the Copula function describes the correlation between multivariate random variables. Its main idea is to decompose the joint distribution function of the multivariate variables into multiple marginal distribution functions, and then connect them using a Copula function. According to Sklar's theorem, if the random variables... The marginal cumulative distribution function CDF is If the joint distribution function is G, then there must exist a copula function C such that: （15） In the above formula, For each daily load characteristic index, there are random variables, specifically the daily average load, peak load, peak-to-valley difference, peak load rate, and valley load rate. The corresponding marginal cumulative distribution function; the joint probability density function is the combination of the various marginal cumulative distribution functions using the Copula function, as shown in the above formula. Random sampling refers to randomly selecting multiple sets of daily load characteristic indicators, with each set of daily load characteristic indicators representing the daily load characteristic indicators of a certain day. In step (3-3), by statistically analyzing the historical load characteristic indicators, it was found that the daily average load, daily peak load, and daily peak-to-valley load difference show a significant upward trend over time, while the peak load rate and valley load rate show no obvious regularity. Therefore, the median changes of these three load characteristic indicators for the target year are generated using a multivariate grey system model. A normal distribution with median deviation as the mean and standard deviation of 1 is constructed, thus obtaining the load characteristic indicators for the target year. The distribution of load characteristic indicator X is expressed as follows: （16） in, The median deviation is the difference between the median of the target year's load characteristic index and the median of the typical year. Let the variance be, and ; In steps (3-4), the optimized model is as follows: （17） In the formula: L(t) is the generated daily load sequence; The selected baseline load sequence; The basic idea behind optimizing model constraints is that the daily characteristic indices of the generated sequence are consistent with the daily characteristic indices obtained from sampling, that is: In the formula, These correspond to the daily average load, daily peak load, daily peak-to-valley load difference, daily peak load rate, and daily valley load rate, respectively. and These represent the times when the maximum and minimum loads occur in the baseline load curve, respectively. and These represent the peak load period and the number of peak load periods, respectively. and These represent the periods of load trough and the number of periods of load trough, respectively. The above optimization model is a quadratic programming model. Solving this optimization problem will yield the optimal daily load sequence that meets the load index requirements.

8. A system for generating medium- to long-term provincial load time-series scenarios driven by meteorological data and taking into account accumulated temperature effects, characterized in that: The method for generating medium- and long-term provincial load time-series scenarios driven by meteorological data and taking into account accumulated temperature effects, as described in claim 1, includes: The analysis module is used to perform correlation analysis between load and meteorological factors, and to divide the annual load sequence into meteorologically sensitive season load and non-meteorologically sensitive season load; The module for generating time-series load scenarios for meteorologically sensitive seasons is used to decompose the load for meteorologically sensitive seasons, construct a temperature cumulative effect model, a multivariate grey system model, and a CNN-BiLSTM-Attention model, and generate time-series load scenarios for meteorologically sensitive seasons. The non-meteorologically sensitive season load time series scenario generation module is used to select daily load characteristic indicators for non-meteorologically sensitive season loads, perform load indicator probability distribution fitting and sampling, obtain non-meteorologically sensitive season load characteristic indicators for the target year, and generate non-meteorologically sensitive season load time series scenarios based on daily characteristic indicators. The target year load time series scenario generation module is used to combine the load of meteorologically sensitive seasons with the load of non-meteorologically sensitive seasons to obtain the target year load time series scenario.

Images