Power prediction method and system considering net load power and meteorological factors

Abstract

The invention discloses a power prediction method and system considering net load power and meteorological factors, and the method comprises the steps: obtaining historical net load data and corresponding meteorological data in a preset time period, and carrying out the preprocessing, and obtaining net load time sequence data and meteorological time sequence data; based on the net load time sequence data, carrying out stability test and differential processing to obtain stable net load time sequence data; establishing an optimal SARIMA model based on the stationary net load time sequence data; extracting trend features and seasonal features based on the optimal SARIMA model; forming a training set based on the meteorological time sequence data, the trend features and the seasonal features; constructing a power prediction model, and training based on the training set and the loss function to obtain a trained power prediction model; and obtaining future meteorological time series data, historical trend characteristics and historical seasonal characteristics of the target time period, and inputting the data, the historical trend characteristics and the historical seasonal characteristics into the trained power prediction model to obtain a net load power prediction value. And the accuracy and the reliability of power prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power prediction, and more particularly to a power prediction method and system taking into account net load power and meteorological factors. Background Art

[0002] Currently, against the backdrop of increasingly complex energy production and consumption, integrated source-load power forecasting, as a key technology for power system optimization and dispatch, has become a research hotspot. Traditional load forecasting relies primarily on user electricity usage data. However, with the large-scale integration of renewable energy sources (such as wind and photovoltaic power) into the power grid, the supply and demand balance in the power system has become even more complex. Renewable energy sources such as wind and photovoltaic power are highly volatile and uncertain, and grid load is significantly affected by factors such as weather and seasonal variations.

[0003] Existing prediction methods usually ignore the power situation on the renewable energy power generation side, or only predict the power consumption on the power user side, failing to effectively reflect the relationship and dynamic changes between the two, resulting in low accuracy and reliability of power prediction.

[0004] Therefore, how to improve the accuracy and reliability of power prediction is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0005] In view of this, the present invention provides a power forecasting method and system taking into account net load power and meteorological factors, thereby improving the accuracy and reliability of power forecasting.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A power forecasting method taking into account net load power and meteorological factors, comprising:

[0008] Obtain historical net load data and corresponding meteorological data for a preset period and preprocess them to obtain net load time series data and meteorological time series data;

[0009] Performing a stationarity test and differential processing based on the net load time series data to obtain stable net load time series data;

[0010] Establishing an optimal SARIMA model based on the steady net load time series data;

[0011] Extracting trend characteristics and seasonal characteristics based on the optimal SARIMA model;

[0012] A training set is formed based on the meteorological time series data, the trend characteristics and the seasonal characteristics;

[0013] Constructing a power prediction model and training it based on the training set and the loss function to obtain a trained power prediction model;

[0014] Obtain future meteorological time series data, historical trend characteristics, and historical seasonal characteristics for the target period and input them into the trained power forecasting model to obtain a net load power forecast value.

[0015] Preferably, the pretreatment includes:

[0016] Based on the historical net load data and the meteorological data, missing value processing, outlier processing, timestamp processing, data standardization and data time resolution unification are performed respectively and in sequence;

[0017] The missing value treatment is different based on the amount of missing data:

[0018] If the amount of missing data is less than or equal to a first threshold, the missing value is filled using the average of the previous and next data points or other interpolation methods;

[0019] If the amount of missing data is greater than a first threshold and less than or equal to a second threshold, the missing values ​​are filled with zero or the mean;

[0020] If the amount of missing data is greater than a second threshold, the record containing the missing value is directly deleted.

[0021] Preferably, the unified data time resolution specifically includes:

[0022] The historical net load data and meteorological data after the missing value processing, the outlier processing, the timestamp processing, and the data standardization processing are recorded as standardized net load data and standardized meteorological data, respectively;

[0023] Correspondingly determining a time resolution of the payload data and a time resolution of the meteorological data based on the standardized payload data and the standardized meteorological data;

[0024] comparing the time resolution of the payload data and the time resolution of the meteorological data with the target resolution respectively;

[0025] If it is greater than the target resolution, the high-frequency data is converted into low-frequency data using a summarization method;

[0026] If it is less than or equal to the target resolution, no processing is performed;

[0027] Based on the above resolution discrimination processing, pre-processed net load data and pre-processed meteorological data are obtained;

[0028] The pre-processed net load data and the pre-processed meteorological data are re-indexed based on the time index to ensure that the timestamps of all data sources are aligned, and the net load time series data and the meteorological time series data are obtained accordingly.

[0029] Preferably, the method for acquiring steady net load time series data is:

[0030] Constructing a test regression model based on the net load time series data;

[0031] Estimate the parameter items of the test regression model based on the least squares method to obtain parameter item estimates;

[0032] Obtaining residual variance based on the estimated values ​​of the parameter terms;

[0033] Obtaining a corrected standard error based on the residual variance;

[0034] Obtaining a test statistic based on the parameter estimate and the standard error;

[0035] determining a critical value, and determining whether the test statistic is less than the critical value;

[0036] If so, it is determined that the net load time series data has a unit root and is non-stationary data;

[0037] Otherwise, determining that the net load time series data is stable data and using it as the stable net load time series data;

[0038] Differential processing is performed based on the non-stationary net load time series data to obtain differential load data, and the above-mentioned stationarity test is cyclically performed based on the differential load data until the stable net load time series data is obtained.

[0039] Preferably, establishing the optimal SARIMA model specifically includes:

[0040] Establishing an initial SARIMA model based on the steady net load time series data;

[0041] Obtaining an optimal model order of the initial SARIMA model based on the steady net load time series data;

[0042] Substituting the optimal model order into the initial SARIMA model to obtain a general SARIMA model;

[0043] Perform maximum likelihood estimation based on the general SARIMA model to obtain optimal model parameters;

[0044] Substituting the optimal model parameters into the general SARIMA model, the optimal SARIMA model is obtained.

[0045] Preferably, the initial SARIMA model is specifically:

[0046] Non-seasonal part:

[0047] (1-φ1B-φ2B 2 -...-φ p B p )(1-B) d Y t =θ0+(θ1B+θ2B 2 +...+θ q B q )ε t ;

[0048] Seasonal section:

[0049] (1-Φ1B m -Φ2B 2m -...-Φ P B Pm )(1-B m ) D Y t =Θ0+(Θ1B m +Θ2B 2m +...+Θ Q B Qm )ε t ;

[0050] Among them, φ p represents the coefficient of the pth non-seasonal autoregressive part, p represents the order of non-seasonal autoregressive, B represents the lag term coefficient, d represents the number of non-seasonal differences, Y t represents the value of the steady net load time series data at time t, θ q represents the coefficient of the qth non-seasonal moving average part, q represents the order of the non-seasonal moving average part, ε t represents the error term, Φ p represents the coefficient of the pth non-seasonal autoregressive part, m represents the seasonal period, D represents the number of seasonal differences, Θ Q represents the coefficient of the seasonal moving average part, and Q represents the order of the seasonal moving average part.

[0051] Preferably, the method for obtaining the optimal model order is:

[0052] Obtaining a seasonal autocorrelation graph, a seasonal partial autocorrelation graph, a non-seasonal autocorrelation graph and a non-seasonal partial autocorrelation graph based on the steady load time series data;

[0053] Obtaining a candidate order value of a seasonal sliding average and a candidate order value of a seasonal autoregression based on the seasonal autocorrelation graph and the seasonal partial autocorrelation graph;

[0054] Obtaining corresponding order candidate values ​​of the non-seasonal sliding average and order candidate values ​​of the non-seasonal autoregressive based on the non-seasonal autocorrelation graph and the non-seasonal partial autocorrelation graph;

[0055] Traversing all the candidate order values ​​based on a step-by-step grid search, fitting the initial SARIMA model to obtain all intermediate SARIMA models;

[0056] The BLC values ​​are calculated based on all the intermediate SARIMA models, and the model order of the intermediate SARIMA model corresponding to the minimum BLC value is selected as the optimal model order.

[0057] Preferably, the method for obtaining the optimal model parameters is:

[0058] Based on the general SARIMA model, the initial parameters are obtained using the least squares method;

[0059] Obtaining residual values ​​of the general SARIMA model based on the initial parameters;

[0060] constructing a likelihood function for a single observation based on the residual value;

[0061] Obtaining a joint likelihood function based on all of the likelihood functions;

[0062] Obtaining a log-likelihood function based on the joint likelihood function;

[0063] The gradient of the log-likelihood function with respect to each model parameter is calculated based on the gradient descent method, and the model parameters are gradually updated along the gradient descent direction until the log-likelihood function converges to a maximum value, and the finally updated model parameters are used as the optimal model parameters.

[0064] Preferably, obtaining a trained power prediction model specifically includes:

[0065] Iteratively training the power prediction model based on the training set:

[0066] The power prediction model calculates a current deviation based on a current predicted value and a true value in each round of iterative training;

[0067] Training the power prediction model by introducing the current deviation as a new feature into a training set to obtain a new training set;

[0068] The power prediction model fits the structure of the current deviation adjustment tree in each round of iterative training, and finally obtains the power prediction model that minimizes the loss function as the trained power prediction model.

[0069] A power prediction system taking into account net load power and meteorological factors, comprising: a data processing module, a stability test module, a feature extraction module, a training set acquisition module, a model training module and a result output module;

[0070] The data processing module is used to obtain historical net load data and corresponding meteorological data for a preset period and perform preprocessing to obtain net load time series data and meteorological time series data;

[0071] The stability test module is used to perform stability test and differential processing based on the net load time series data to obtain stable net load time series data;

[0072] The feature extraction module is used to establish an optimal SARIMA model based on the steady net load time series data; and extract trend features and seasonal features based on the optimal SARIMA model;

[0073] The training set acquisition module is used to form a training set based on the meteorological time series data, the trend characteristics and the seasonal characteristics;

[0074] The model training module is used to construct a power prediction model and perform training based on the training set and the loss function to obtain a trained power prediction model;

[0075] The result output module is used to obtain future meteorological time series data, historical trend characteristics and historical seasonal characteristics of the target period and input them into the trained power prediction model to obtain the net load power prediction value.

[0076] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a power forecasting method and system that takes into account net load power and meteorological factors, which comprehensively considers the coupling relationship between the power supply side and the demand side in the power system and the influence of meteorological factors. Compared with the traditional forecasting method that considers the demand side load or the power of the power generation side alone, the present invention takes the difference between the power of the power supply side and the demand side, that is, the net load power, as the core forecasting target, aiming to accurately capture the dynamic interaction effect between the two; the present invention also introduces meteorological elements, combines the time series features extracted by the time series model, and introduces the LightGBM model to fully explore the linear and nonlinear relationship between historical load data and environmental factors, thereby achieving more accurate load forecasting; compared with the traditional single-side forecasting method, the method of the present invention can more effectively reflect the inherent coupling mechanism between the power supply and demand sides, while considering the impact of external meteorological changes on power fluctuations, thereby improving the forecast accuracy and stability; the present invention can provide accurate load forecasting and decision support for power system scheduling, demand response, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0078] Figure 1 This is a flow chart of a power forecasting method taking into account net load power and meteorological factors provided by the present invention.

[0079] Figure 2 This is a flow chart of the method for acquiring steady net load time series data provided by the present invention.

[0080] Figure 3 This is a schematic diagram comparing the predicted value and actual value curves of the power prediction model provided by the present invention.

[0081] Figure 4 This is a schematic diagram comparing the predicted values ​​and actual values ​​of different power prediction models provided by the present invention.

[0082] Figure 5 This is a schematic diagram of the structure of a power prediction system taking into account net load power and meteorological factors provided by the present invention. DETAILED DESCRIPTION

[0083] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0084] Example 1

[0085] like Figure 1 As shown, an embodiment of the present invention discloses a power forecasting method taking into account net load power and meteorological factors, including:

[0086] Obtain historical net load data and corresponding meteorological data for a preset period and preprocess them to obtain net load time series data and meteorological time series data;

[0087] Based on the net load time series data, stationarity test and differential processing are performed to obtain stable net load time series data;

[0088] Establish an optimal SARIMA model based on steady net load time series data;

[0089] Trend characteristics and seasonal characteristics are extracted based on the optimal SARIMA model;

[0090] The training set is composed of meteorological time series data, trend characteristics and seasonal characteristics;

[0091] Construct a power prediction model and train it based on the training set and loss function to obtain a trained power prediction model;

[0092] Obtain future meteorological time series data, historical trend characteristics, and historical seasonal characteristics for the target period and input them into the trained power forecasting model to obtain the net load power forecast value.

[0093] Example 2

[0094] An embodiment of the present invention discloses a power forecasting method taking into account net load power and meteorological factors, comprising:

[0095] The historical net load data and corresponding meteorological data of the preset time period are obtained and preprocessed to obtain the net load time series data and meteorological time series data.

[0096] Preferably, the net load data, ie, the net load power, is obtained based on the difference between the household side load data and the new energy power generation data.

[0097] Preferably, meteorological data corresponding to the preset time period and the historical net load data is obtained; in this embodiment, the meteorological data includes: temperature, humidity, wind speed, wind direction and sunshine radiation value.

[0098] Preferably, the pretreatment comprises:

[0099] Based on the historical net load data and meteorological data, missing value processing, outlier processing, timestamp processing, data standardization and data time resolution unification are carried out in sequence to obtain net load time series data and meteorological time series data;

[0100] Missing value handling is done differently based on the amount of missing data:

[0101] If the amount of missing data is less than or equal to the first threshold, the missing value is filled using the average of the previous and next data points or other interpolation methods;

[0102] If the amount of missing data is greater than the first threshold and less than or equal to the second threshold, the missing values ​​are filled with zero or the mean;

[0103] If the amount of missing data is greater than the second threshold, the records containing missing values ​​are directly deleted.

[0104] Preferably, for small, sparsely distributed missing values ​​in continuous time series data, linear interpolation or temporal interpolation is preferred for estimation to preserve data trends. For cases with a high proportion of missing values ​​or a random missing pattern, the mean, median, or historical distribution pattern is used for reasonable filling. When missing values ​​are severe or the missing data are not valuable for reconstruction, the corresponding data samples are directly removed to ensure overall data quality. The above methods can be combined according to specific application scenarios, with flexibility and adaptability, improving the accuracy and robustness of subsequent model training and prediction.

[0105] Outliers may be the result of data entry errors or extreme events. In this embodiment, the boxplot method is used as a method for handling outliers. The boxplot method is a simple and effective outlier detection method. Its main features include being insensitive to data distribution, simple and intuitive, and able to analyze data series with different units and dimensions. It can also be used to perform outlier detection on multiple variables or data sets in batches. It is not affected by data distribution and is easy to understand and apply.

[0106] Preferably, the box plot mainly uses the quartiles of the data to identify outliers, which are defined as follows:

[0107] ① First quartile Q1: 25% of the values ​​in the data are below this value;

[0108] ②The third quartile Q3: 75% of the values ​​in the data are below this value;

[0109] The first quartile (Q1) refers to the value at the 25th percentile after sorting the sample data from smallest to largest, reflecting the distribution characteristics of the lower quarter of the data; the third quartile (Q3) refers to the value at the 75th percentile after sorting, reflecting the distribution characteristics of the upper quarter of the data;

[0110] The process for determining Q1 and Q3 involves first sorting the sample data from smallest to largest. After sorting, Q1 is determined by the median of the subset to the left of the median (Q2), while Q3 is determined by the median of the subset to the right of the median. For an odd number of data, the median itself is excluded and the subsets are taken separately. For an even number of data, the front and back halves are directly used as subsets for calculation. For example, suppose the dataset is [1, 3, 5, 7, 9, 11, 13] (an odd number of elements), the median Q2 = 7, the left subset is [1, 3, 5], and the right subset is [9, 11, 13], so Q1 = 3 and Q3 = 11. Similarly, for the data set [1,3,5,7,9,11] (even number of elements), the median Q2 = (5+7) / 2 = 6, the left subset is [1,3], and the right subset is [9,11], so Q1 = (1+3) / 2 = 2, Q3 = (9+11) / 2 = 10;

[0111] ③Interquartile range IQR: its value is equal to Q3-Q1;

[0112] ④ Abnormal value: value exceeding Q3+1.5*IQR (upper whisker) or lower than Q1-1.5*IQR (lower whisker);

[0113] When using the boxplot method to identify outliers, the first quartile, third quartile, and interquartile range are calculated to determine the values ​​of the upper and lower whiskers. Any value smaller than the lower whisker or larger than the upper whisker is considered an outlier. These outliers are removed to optimize data accuracy and improve prediction accuracy.

[0114] Preferably, missing value processing and outlier processing are used together as data cleaning processing to ensure the quality and availability of the original data and lay a good foundation for subsequent modeling and analysis.

[0115] Preferably, timestamp processing involves parsing and formatting the timestamp fields in historical payload and meteorological data before data standardization. By standardizing the time format and extracting key time features, data consistency and integrity across the time dimension are ensured. This step provides an accurate time foundation for subsequent data standardization and model input feature construction.

[0116] Load data from historical payload and weather data, converting timestamp fields from strings to datetime objects for time series analysis.

[0117] Preferably, data standardization specifically includes:

[0118] Data standardization is performed on the basis of timestamp processing. Data standardization mainly converts the data into a format suitable for model processing to improve the stability and prediction performance of the model, while eliminating the impact of different scales on model training. The present invention adopts Z-Score standardization to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0119] Preferably, the unified data time resolution specifically includes:

[0120] The historical net load data and meteorological data after missing value processing, outlier processing, timestamp processing, and data standardization are recorded as standardized net load data and standardized meteorological data respectively;

[0121] Determining the time resolution of the net load data and the time resolution of the meteorological data based on the standardized net load data and the standardized meteorological data;

[0122] The time resolution of the net load data and the time resolution of the meteorological data are compared with the target resolution respectively;

[0123] If it is larger than the target resolution, the high-frequency data is converted into low-frequency data using a summarization method;

[0124] If it is less than or equal to the target resolution, no processing is performed to ensure that no data is missed;

[0125] Based on the above resolution discrimination processing, pre-processed net load data and pre-processed meteorological data are obtained;

[0126] The pre-processed net load data and pre-processed meteorological data are re-indexed based on the time index to ensure that the timestamps of all data sources are aligned, and the corresponding net load time series data and meteorological time series data are obtained.

[0127] Preferably, the specific target resolution is based on the actual forecasting task requirements. Based on the above processing, the standardized net load data and the standardized meteorological data are unified into a common time resolution to ensure the temporal consistency of the data.

[0128] Preferably, net load time series graphs and meteorological time series graphs are drawn based on the net load time series data and meteorological time series data respectively to observe the changes in the data at different time points and whether they show significant trends, seasonality or cyclical fluctuations. By analyzing the net load time series graph, obvious seasonal and trend components are identified, verifying the feasibility of using the SARIMA model to extract trend and seasonal features; through the joint time series graph of meteorology and load, the synchronous fluctuation trend of load and external factors such as temperature and humidity is revealed, providing data support for the subsequent incorporation of meteorological data into the LightGBM model as feature input.

[0129] Based on the net load time series data, stationarity test and differential processing are performed to obtain stable net load time series data.

[0130] Preferably, Figure 2 As shown in Figure 2, the method for obtaining steady net load time series data is:

[0131] Construct a test regression model based on the net load time series data;

[0132] The parameter items of the regression model are estimated based on the least squares method to obtain the estimated values ​​of the parameter items;

[0133] Obtain the residual variance based on the estimated values ​​of the parameter terms;

[0134] The corrected standard error is obtained based on the residual variance;

[0135] The test statistic is obtained based on the parameter estimates and standard errors;

[0136] Determine the critical value and determine whether the test statistic is less than the critical value;

[0137] If so, it is determined that the net load time series data has a unit root and is non-stationary data;

[0138] Otherwise, the net load time series data is determined to be stationary data and is regarded as stationary net load time series data;

[0139] Perform differential processing based on non-stationary net load time series data to obtain differential load data;

[0140] The above-mentioned stationarity test is cyclically performed based on the differential load data until the stable net load time series data is obtained.

[0141] Preferably, this embodiment uses the Phillips-Perron test (hereinafter referred to as the PP test) to check the stationarity of the net load time series data. The PP test is similar to the ADF test and also tests whether a time series has a unit root. However, the PP test is more robust to autocorrelation and heteroscedasticity and is more stable than the ADF test when dealing with high-order autocorrelation or heteroscedasticity in time series. The PP test corrects heteroscedasticity and autocorrelation in the data by adjusting the standard error in the ADF test, making it more suitable for load data with seasonal fluctuations or strong volatility.

[0142] In time series analysis, a unit root refers to a situation in which the autoregressive coefficient in a time series model is equal to 1. A unit root indicates that the characteristics of the series may change significantly over time, i.e., the series is non-stationary. Specifically, for an autoregressive (AR) model, if the autoregressive coefficient is equal to 1, it indicates the presence of a unit root. This situation causes the mean and variance of the data to change over time. Here, the PP test is used to test whether the model has a unit root, and differencing is used to ensure data stationarity.

[0143] Preferably, the test regression model of this embodiment adopts a test regression model with a constant term, a time trend term, and a lag term, and its specific model is as follows:

[0144] y t =α+βt+γy t-1 +ζ t ;

[0145] y t represents the net load time series data to be tested, α represents the constant term, βt represents the time trend term, γy t-1 Represents the autoregressive term, which indicates the impact of the previous period's value on the current value and is used to detect unit roots, ζ t represents the error term.

[0146] Preferably, the parameter items of the test regression model are estimated based on the least squares method to obtain the estimated values ​​of the parameter items:

[0147] The residual sum of squares S is obtained based on the predicted values ​​of the test regression model:

[0148]

[0149] Among them, x t represents the actual observation value, (α+γy t-1 +βt) represents the predicted value of the test regression model, and T represents the number of samples;

[0150] Estimate the parameters α, β, and γ by minimizing the residual sum of squares:

[0151]

[0152] in, are the estimated values ​​of parameters α, β, and γ respectively;

[0153] For the partial derivative of α:

[0154] Set it to zero, and do the same for β and γ in turn, to obtain a set of linear equations about α, β, and γ, which can be solved by matrix operations:

[0155]

[0156] Among them, X contains the independent variable (i.e., y t-1 and t), y represents the matrix containing the dependent variable x t vector.

[0157] Preferably, based on the estimated value of the parameter and Get the residual variance σ 2 :

[0158]

[0159] in, Represents the residual.

[0160] Preferably, based on the residual variance σ 2 Get the corrected standard error

[0161]

[0162] Preferably, based on the estimated value of the parameter and standard error Get the test statistic t γ :

[0163]

[0164] Preferably, the corresponding critical value is obtained by consulting the Phillips-Perron critical value table.

[0165] Preferably, for non-stationary data, differential processing is performed:

[0166] k' t =k t -k t-1 ;

[0167] k t represents the non-stationary net load time series data at time t, k t-1 represents the non-stationary net load time series data at time t-1, k' t It is the differential load data after one difference.

[0168] The optimal SARIMA model is established based on the steady net load time series data.

[0169] Preferably, the SARIMA model combines autoregression (AR) and moving average (MA) methods to capture the changing trend in the time series for forecasting. The SARIMA model mainly combines autoregression (AR), differencing (I), moving average (MA) and seasonal components.

[0170] Preferably, the autoregressive part (AR) uses the past values ​​of the stationary net load time series data itself to predict the current value. Autoregression means that there is a certain linear relationship between the current value and its values ​​at several previous moments. The order p of the autoregressive part indicates how many past time points are used to make the prediction. Its expression is:

[0171] Y t =φ1Y t-1 +φ2Y t-2 +...+φ p Y t-p +δ t ;

[0172] Y t Represents the value of the steady net load time series data at time t, φ1...,φ p is the autoregressive coefficient, δ t Represents the white noise error term, which is an unpredictable random disturbance;

[0173] The differencing part (I) is used to convert a non-stationary time series into a stationary one. A stationary time series has a constant mean and variance over time, making it suitable for modeling and forecasting. Differencing is performed by calculating the difference between adjacent moments. If the series is still non-stationary after one differencing operation, a second or higher order differencing operation can be performed. The differencing order d indicates how many differencing operations are required to make the series stationary.

[0174] The moving average (MA) uses the past forecast errors of the time series to predict the current value. A moving average model implies a linear relationship between the current value and the forecast errors at several previous moments. The order (q) of the moving average model indicates how many past errors are used to make the forecast; its expression is:

[0175] Y t =λ t +θ1λ t-1 +θ2λ t-2 +...+θ q λ t-q ;

[0176] Among them, Y t represents the value of the steady net load time series data at time t, λ t ,λ t-1 ,...,λ t-q Denotes the prediction error, θ1, θ2, ..., θ q Represents the moving average coefficient.

[0177] Preferably, the autoregressive component (AR) in the SARIMA model is used to describe the linear correlation between the current net load power value and its historical value; the differential component (I) performs differential processing on the original net load time series data to eliminate the trend component in the series and enhance its stability; the moving average component (MA) is used to model the relationship between the current value and the historical error term to fit the impact of random disturbances. Based on the above three basic components, the seasonal parameter term is introduced to further construct the SARIMA model, which is used to simultaneously model the long-term trend, short-term volatility, and seasonal periodicity characteristics of the time series, thereby accurately characterizing the evolution law of the net load power series and extracting multidimensional features.

[0178] Preferably, establishing the optimal SARIMA model specifically includes:

[0179] Establish an initial SARIMA model based on the steady net load time series data;

[0180] The optimal model order of the initial SARIMA model is obtained based on the steady net load time series data;

[0181] Substitute the optimal model order into the initial SARIMA model to obtain the general SARIMA model;

[0182] Based on the general SARIMA model, the maximum likelihood estimation is performed to obtain the optimal model parameters;

[0183] Based on the optimal model parameters, the general SARIMA model is substituted to obtain the optimal SARIMA model.

[0184] Preferably, the initial SARIMA model is specifically:

[0185] Non-seasonal part:

[0186] (1-φ1B-φ2B 2 -...-φ p B p )(1-B) d Y t =θ0+(θ1B+θ2B 2 +...+θ q B q )ε t ;

[0187] Seasonal section:

[0188] (1-Φ1B m -Φ2B 2m -...-Φ P B Pm )(1-B m ) D Y t =Θ0+(Θ1B m +Θ2B 2m +...+Θ Q B Qm )ε t ;

[0189] Among them, φ p represents the coefficient of the pth non-seasonal autoregressive part, p represents the order of non-seasonal autoregressive, B represents the lag term coefficient, d represents the number of non-seasonal differences, Y t represents the value of the steady net load time series data at time t, θ q represents the coefficient of the qth non-seasonal moving average part, q represents the order of the non-seasonal moving average part, ε t represents the error term, Φ p represents the coefficient of the pth non-seasonal autoregressive part, m represents the seasonal period, D represents the number of seasonal differences, Θ Q represents the coefficient of the seasonal moving average part, and Q represents the order of the seasonal moving average part.

[0190] Preferably, the SARIMA model adds modeling of seasonal fluctuations on the basis of ARIMA, which can better handle the cyclical changes of time series data and make the time series stable through differential operations, which is conducive to model fitting and prediction.

[0191] Preferably, the method for obtaining the optimal model order is:

[0192] Based on the steady load time series data, seasonal autocorrelation diagram, seasonal partial autocorrelation diagram, non-seasonal autocorrelation diagram and non-seasonal partial autocorrelation diagram are obtained;

[0193] Based on the seasonal autocorrelation diagram and the seasonal partial autocorrelation diagram, the order candidate values ​​of the seasonal moving average and the order candidate values ​​of the seasonal autoregression are obtained;

[0194] Based on the non-seasonal autocorrelation graph and the non-seasonal partial autocorrelation graph, the order candidate values ​​of the non-seasonal sliding average and the order candidate values ​​of the non-seasonal autoregressive are obtained;

[0195] Based on the step-by-step grid search, all the candidate order values ​​mentioned above are traversed, the initial SARIMA model is fitted, and all the intermediate SARIMA models are obtained;

[0196] The BLC values ​​are calculated based on all the intermediate SARIMA models and the model order of the intermediate SARIMA model corresponding to the minimum BLC value is selected as the optimal model order.

[0197] Preferably, observe the non-seasonal partial autocorrelations and find the locations of significant truncation. If the non-seasonal partial autocorrelations decay sharply after a certain lag order, then that order is a candidate for the order of the non-seasonal autoregressive. Observe the non-seasonal autocorrelations and find the locations of significant truncation. If the non-seasonal autocorrelations decay sharply after a certain lag order, then that order is a candidate for the order of the non-seasonal moving average.

[0198] Observe the seasonal partial autocorrelogram and find the location of significant truncation. If the seasonal partial autocorrelogram decays sharply after a certain lag order, then that order is a candidate for the seasonal autoregressive order. Observe the seasonal autocorrelogram and find the location of significant truncation. If the seasonal autocorrelogram decays sharply after a certain lag order, then that order is a candidate for the seasonal moving average order.

[0199] Preferably, the Bayesian Information Criterion (BIC) formula is:

[0200]

[0201] Among them, n represents the sample size, that is, the number of observation data points of the steady net load time series data, It represents the estimated variance of the model residuals, that is, the variance of the residuals after model fitting. M represents the number of estimated parameters in the model. For the SARIMA model, it includes the seasonal difference order, non-seasonal difference order, AR (autoregressive) order, MA (moving average) order, as well as parameters such as seasonal autoregressive (SAR), seasonal moving average (SMA) and seasonal difference.

[0202] Preferably, the method for obtaining the optimal model parameters is:

[0203] The initial parameters are obtained using the least squares method based on the general SARIMA model;

[0204] Obtain the residual value of the general SARIMA model based on the initial parameters;

[0205] Construct a likelihood function for a single observation based on the residual values;

[0206] Based on all likelihood functions, a joint likelihood function is obtained;

[0207] The log-likelihood function is obtained based on the joint likelihood function;

[0208] The gradient of the log-likelihood function with respect to each model parameter is calculated based on the gradient descent method, and the model parameters are gradually updated along the gradient descent direction until the log-likelihood function converges to the maximum value. The finally updated model parameters are used as the optimal model parameters.

[0209] Preferably, the residual value where y t is the true value, is the predicted value, based on the current parameter estimates.

[0210] Preferably, the likelihood function for a single observation is:

[0211]

[0212] Among them, ψ 2 represents the variance of the error term, represents the constant part of the normal distribution, represents the exponential part of each error term.

[0213] Preferably, the joint likelihood function is specifically:

[0214]

[0215] Here, n represents the number of likelihood functions for a single observation.

[0216] Preferably, the logarithm of the joint likelihood function is taken to obtain the logarithmic likelihood function:

[0217]

[0218] Preferably, the gradient calculation formula is:

[0219] Update parameters:

[0220] Among them, θ i represents the model parameters, represents the updated model parameters, Represents the model parameters before update, α represents the learning rate, which determines the step size of each iterative update.

[0221] Preferably, the log-likelihood function converges to a maximum value, and the goal of the optimization process is:

[0222]

[0223] Preferably, the final model parameters include:

[0224] The coefficients of the AR part φ1, φ2, ..., φ p ;

[0225] The coefficients of the MA part θ1, θ2, ..., θ q ;

[0226] The coefficients of the SAR part Φ1, Φ2, ..., Φ p ;

[0227] The coefficients Θ1, Θ2, ..., Θ of the SMA part Q ;

[0228] The variance of the error term ψ 2 .

[0229] Trend characteristics and seasonal characteristics are extracted based on the optimal SARIMA model.

[0230] Preferably, the trend term is captured by the non-seasonal part, and the autoregressive part in the non-seasonal part formula of the optimal SARIMA model is fitted to obtain the trend feature for extraction:

[0231]

[0232] Among them, Y t trend represents the trend characteristic, that is, the long-term trend after removing seasonal fluctuations, μ represents the constant term, which represents the long-term average level of the sequence; φ i represents the coefficient of the autoregressive part, capturing the impact of past time points (lags); Y t-i is the observed value at the ith lag moment in the stationary net load time series data.

[0233] Preferably, seasonal characteristics reflect the cyclical fluctuations in the data that occur with seasonal changes. In the SARIMA model, seasonal characteristics are generally captured by the seasonal autoregressive (SAR) and seasonal moving average (SMA) parts. The seasonal effect is removed by seasonally differencing the data ((1-BS)D), thereby obtaining a time series without seasonal fluctuations. This method uses the difference between the original series and the trend series to solve the seasonal characteristics, avoiding the complex seasonal decomposition process, reducing the dependence on the seasonal model, and simplifying the computational complexity of the model. The formula is as follows:

[0234] Y t saesonal =Y t -Y t trend ;

[0235] Y t saesonal represents seasonal characteristics, Y t Represents steady net load time series data; seasonal characteristics reflect the periodic fluctuations in the data and remove the influence of long-term trends.

[0236] The training set is composed of meteorological time series data, trend characteristics and seasonal characteristics.

[0237] Preferably, the training set is a feature matrix of training data, specifically expressed as Xl:

[0238] Xl=[temp,humid,WindSpeed,WindDirection,ssrd,sarima_trend,sarima_seasonal];

[0239] Among them, temp represents temperature; humid represents humidity; WindSpeed ​​represents wind speed; WindDirection represents wind direction; ssrd represents solar radiation value; sarima_trend represents trend characteristics; and sarima_seasonal represents seasonal characteristics.

[0240] A power prediction model is constructed and trained based on the training set and loss function to obtain a trained power prediction model.

[0241] Preferably, the power prediction model in this embodiment adopts the LightGBM model. The LightGBM model is based on a framework built on the Gradient Boosting Decision Tree (GBDT). The goal is to gradually approach the true label value by continuously building a tree model.

[0242] Preferably, before training the power prediction model, the features in the training set are normalized. Normalization can eliminate the impact of different feature scale differences on model training, making the contribution of each feature to the model more balanced. The normalization formula is:

[0243]

[0244] Where h is the mean of the feature column and v is the standard deviation of the feature column. After standardization, the mean of all features is 0 and the standard deviation is 1.

[0245] Preferably, obtaining a trained power prediction model specifically includes:

[0246] Iteratively train the power prediction model based on the training set:

[0247] The power prediction model calculates the current deviation based on the current predicted value and the true value in each round of iterative training;

[0248] The power prediction model is trained by introducing the current deviation as a new feature into the training set to obtain a new training set;

[0249] The power prediction model fits the structure of the current bias adjustment tree in each round of iterative training, and finally obtains the power prediction model that minimizes the loss function as the trained power prediction model.

[0250] Preferably, this embodiment selects the mean of historical net load data as the initial prediction value f0(x) of the model:

[0251]

[0252] Among them, f0(x) represents the initial prediction value of the power prediction model, which is the initial prediction value of the power prediction model when there is no tree. The initial prediction value takes the historical net load mean, y i is the i-th historical net load data, n represents the number of historical net load data;

[0253] The model will be based on the current prediction value f m-1 (x) and the true value y to calculate the residual:

[0254] r=yf m-1 (x);

[0255] f m-1 (x) represents the predicted value of the model at the m-1th iteration, which indicates the difference between the current prediction of the model and the true value.

[0256] Introduce the current deviation r as a new feature into the training set to obtain a new training set:

[0257]

[0258] After calculating the residuals, the power prediction model trains a new tree based on the new training set:

[0259]

[0260] in, It represents a decision tree trained based on the input features. The output is a residual correction, which is used to adjust the model's prediction so that the tree predicts a value close to the residual. The prediction value of each tree can be expressed as:

[0261]

[0262] Among them, h m (x) represents the predicted value of the mth tree, λ j is the coefficient of the leaf node, I j (x) is the leaf node indicator function of the tree, K is the number of leaves in the tree;

[0263] For each tree, this method updates the current prediction:

[0264] f m (x) = f m-1 (x)+η·h m (x);

[0265] Among them, f m (x) is the updated predicted value, h m (x) is the predicted value of the current tree, and η is the learning rate, which controls the step size of each update. By adjusting η, the fit of the model can be controlled.

[0266] Each iteration updates the model by calculating the gradient to minimize the loss function

[0267]

[0268] in, is the loss function, f m (x i ) is the predicted value of the i-th sample, y i is the actual value of the i-th sample. The model fits the structure of the residual adjustment tree in each round of iteration and finally finds a model that minimizes the loss function.

[0269] Preferably, in terms of model structure, the present invention has made some changes to the model tree growth strategy and adopted an adaptive tree growth strategy, that is, the traditional tree growth is fixed, starting from the root node and gradually splitting to the leaf node. The adaptive tree growth strategy is to adaptively adjust the depth and splitting strategy of each node based on the importance of the tree nodes at each layer. For example, if certain features have already demonstrated strong predictive capabilities in early nodes, then they can be given priority to split in subsequent nodes, or the role of the feature in subsequent nodes can be limited to avoid overfitting. The model ultimately predicts the short-term net load power time series value for the next 24-72 hours based on actual demand.

[0270] Preferably, the method further includes: evaluating and optimizing based on the trained power prediction model:

[0271] The residual is calculated based on the difference between the predicted value output by the trained power prediction model and the true value. If the trained power prediction model fits the data well, the residual should conform to white noise, that is, the residual should be an independent and identically distributed random variable with zero mean and constant variance.

[0272] White noise test of residuals:

[0273] The residuals are tested for white noise based on their autocorrelation function (ACF) and partial autocorrelation function (PACF). The autocorrelation plot and partial autocorrelation plot are generated based on the ACF and PACF. If the model fits well, the residuals should have no significant autocorrelation, meaning their autocorrelation plot and partial autocorrelation plot should decay rapidly to zero after lag 1. If the autocorrelation coefficients and partial autocorrelation coefficients for all lags in the autocorrelation plot and partial autocorrelation plot are close to zero and within the 95% confidence interval, it can be considered that the residuals have no significant autocorrelation and the model fits well. If the ACF or PACF for some lags deviates significantly from zero (outside the confidence interval), it indicates that the residuals have significant autocorrelation, indicating that the model does not fully fit the data. Wherein Bc represents the critical value of the 95% confidence interval under the standard normal distribution, which is 1.96 in this embodiment, and nr represents the number of observed data points.

[0274] If the autocorrelation plot and partial autocorrelation plot show that there is autocorrelation in the residuals, it means that the model does not fully capture certain patterns in the data, and it is necessary to add AR (autoregressive) or MA (moving average) terms to improve the model.

[0275] If the residual autocorrelation plot and partial autocorrelation plot are not significantly correlated, then the residuals of the model can be considered white noise, indicating that the model has captured the structure of the data well.

[0276] Based on the above ACF and PACF analysis of the residuals, if there is autocorrelation in the residuals, the Q statistic can further verify the existence of autocorrelation. If there is no significant correlation in the residuals, in order to quantitatively analyze whether the residuals of the model are white noise, this method uses the Q statistic (Ljung-Box) to quantitatively test whether the residuals are white noise. The statistical formula is:

[0277]

[0278] Where n is the sample size; m is the number of lag periods; ρ k represents the autocorrelation coefficient at lag k.

[0279] If the p-value (P value) of the Q value is less than the predetermined significance level (for example, 0.05), the null hypothesis is rejected, indicating that there is autocorrelation in the residuals and the model needs to be optimized.

[0280] RMSE calculates the root mean square of the difference between the predicted value and the actual value. The smaller the RMSE, the stronger the predictive ability of the model.

[0281] Parameter optimization:

[0282] By adjusting the key hyperparameters of the model to improve its prediction accuracy and reduce overfitting:

[0283] Determine the initial configuration of the power prediction model to be optimized, including but not limited to the following hyperparameters:

[0284] Learning rate (learning_rate): controls the step size of the model weight update in each iteration, usually in the range of [0,1][0,1][0,1].

[0285] Number of trees (num_trees): The number of trees used in model training.

[0286] Number of leaf nodes (num_leaves): controls the complexity of the tree and is usually set to a smaller value to avoid overfitting.

[0287] Maximum depth (max_depth): Controls the depth of the tree to avoid overfitting of too deep trees.

[0288] Minimum data amount (min_data_in_leaf): Set the minimum number of samples contained in each leaf node to control overfitting.

[0289] Regularization parameters: L1 regularization and L2 regularization, used to control the complexity of the model and prevent overfitting.

[0290] Hyperparameter optimization:

[0291] In order to obtain the optimal hyperparameter combination, this method uses cross-validation as the evaluation criterion. By dividing the dataset into multiple training sets and validation sets, the performance of each hyperparameter combination is evaluated based on the loss function. The selected loss function is the mean square error (MSE):

[0292]

[0293] Y i is the true value, is the predicted value of the model, and n is the number of samples. By minimizing the MSE, the model's prediction accuracy for unknown data is improved.

[0294] Using grid search, we perform an exhaustive search through the candidate value space of predefined hyperparameters, calculate the loss function value corresponding to each hyperparameter combination, and select the hyperparameter combination that minimizes the loss:

[0295]

[0296] Among them, p1, p2, ..., p n is the hyperparameter to be optimized, and Loss is the loss function value after cross-validation.

[0297] The loss function is a criterion for evaluating the prediction accuracy of a model. In this paper, the mean square error is selected as the loss function. By optimizing the model parameters, the prediction error on the training set and the validation set is minimized, that is, the following formula is minimized:

[0298]

[0299] Among them, θ is the model parameter, Y i is the true value, is the predicted value of the model under the parameter θ. By minimizing this loss function, the prediction accuracy of the model can be effectively improved.

[0300] Obtain future meteorological time series data, historical trend characteristics, and historical seasonal characteristics for the target period and input them into the trained power forecasting model to obtain the net load power forecast value.

[0301] Preferably, this embodiment is aimed at short-term prediction (next 24 hours to 72 hours). During the prediction, the meteorological series data for the next 24 hours to 72 hours are obtained, and the trend and seasonal characteristics of the past 24 hours to 72 hours are obtained to form a prediction feature matrix. The prediction feature matrix is ​​input into the trained power prediction model, and the output is the net load power prediction value for the next 24 hours to 72 hours, that is, the net load power value (the time scale is the same as the historical net load data, which is a value of 15 minutes).

[0302] Example 3

[0303] The data collected by this method are the historical net load data from March 14, 2024 to March 14, 2025, and the meteorological data from March 14, 2025 to March 15, 2025. When the historical net load data is accurate and the meteorological data is sufficient, the trained LightGBM model is used to predict the net power situation in the next 24 hours (from 0:00 to 23:45 on March 15, 2025). The predicted results are compared with the actual power and the predictions of common models, and the single-point deviation and accuracy are calculated:

[0304] Single point deviation calculation formula:

[0305]

[0306] Where δ represents the single point deviation, represents the predicted value, and y represents the actual value;

[0307] Accuracy formula:

[0308]

[0309] Among them, P represents the accuracy, n represents the number of data points, and y i Indicates the actual value, represents the predicted value;

[0310] Here, we use the root mean square method of the single-point deviation to obtain the overall deviation and calculate the accuracy, which can better evaluate the accuracy of the model.

[0311] like Figure 3-Figure 4 As shown in the figure, from the perspective of the degree of fit between the model prediction curve and the actual curve, this method has better fit and robustness than other models in terms of overall fit. The SARIMA prediction does not capture the nonlinear characteristics of the data very well, so the overall deviation is large. Although the XGBoost and LgightGBM models capture the overall trend, they still have more deviations than this method. The comparison results of model deviation and accuracy are shown in Table 1:

[0312] Table 1 Comparison of model bias and accuracy

[0313]

[0314] From the results in Table 1, we can more specifically observe that the degree of deviation and overall accuracy of this method have better precision and generalization than common time series models and individual machine learning models.

[0315] This method uses net load power as input data, fully considers the impact of user-side load data and renewable energy power generation data, and considers the impact of meteorological data. It uses the SARIMA model to extract the trend and seasonal characteristics of the data, and comprehensively considers the impact of the data itself and meteorological factors on net load power. At the same time, it uses the more efficient LightGBM model for training and prediction.

[0316] Example 4

[0317] like Figure 5 As shown, a power prediction system taking into account net load power and meteorological factors includes: a data processing module, a stability test module, a feature extraction module, a training set acquisition module, a model training module and a result output module;

[0318] A data processing module is used to obtain historical net load data and corresponding meteorological data for a preset period and perform preprocessing to obtain net load time series data and meteorological time series data;

[0319] The stability test module is used to perform stability test and differential processing based on the net load time series data to obtain stable net load time series data;

[0320] A feature extraction module is used to establish an optimal SARIMA model based on the steady net load time series data; trend features and seasonal features are extracted based on the optimal SARIMA model;

[0321] The training set acquisition module is used to form a training set based on meteorological time series data, trend characteristics and seasonal characteristics;

[0322] A model training module is used to build a power prediction model and perform training based on a training set and a loss function to obtain a trained power prediction model;

[0323] The result output module is used to obtain future meteorological time series data, historical trend characteristics and historical seasonal characteristics of the target period and input them into the trained power prediction model to obtain the net load power prediction value.

[0324] Preferably, the function implementation process of each module in this embodiment corresponds one-to-one to the content of the above method, and will not be repeated here.

[0325] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0326] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A power forecasting method taking into account net load power and meteorological factors, characterized in that: include: Obtain historical net load data and corresponding meteorological data for a preset period and preprocess them to obtain net load time series data and meteorological time series data; Performing a stationarity test and differential processing based on the net load time series data to obtain stable net load time series data; Establishing an optimal SARIMA model based on the steady net load time series data; Extracting trend characteristics and seasonal characteristics based on the optimal SARIMA model; A training set is formed based on the meteorological time series data, the trend characteristics and the seasonal characteristics; Constructing a power prediction model and training it based on the training set and the loss function to obtain a trained power prediction model; Obtain future meteorological time series data, historical trend characteristics, and historical seasonal characteristics for the target period and input them into the trained power forecasting model to obtain a net load power forecast value.

2. A power forecasting method taking into account net load power and meteorological factors according to claim 1, characterized in that: The pretreatment includes: Based on the historical net load data and the meteorological data, missing value processing, outlier processing, timestamp processing, data standardization and data time resolution unification are performed respectively and in sequence; The missing value treatment is different based on the amount of missing data: If the amount of missing data is less than or equal to a first threshold, the missing value is filled using the average of the previous and next data points or other interpolation methods; If the amount of missing data is greater than a first threshold and less than or equal to a second threshold, the missing values ​​are filled with zero or the mean; If the amount of missing data is greater than a second threshold, the record containing the missing value is directly deleted.

3. A power forecasting method taking into account net load power and meteorological factors according to claim 2, characterized in that: The unified data time resolution specifically includes: The historical net load data and meteorological data after the missing value processing, the outlier processing, the timestamp processing, and the data standardization processing are recorded as standardized net load data and standardized meteorological data, respectively; Correspondingly determining a time resolution of the payload data and a time resolution of the meteorological data based on the standardized payload data and the standardized meteorological data; comparing the time resolution of the payload data and the time resolution of the meteorological data with the target resolution respectively; If it is greater than the target resolution, the high-frequency data is converted into low-frequency data using a summarization method; If it is less than or equal to the target resolution, no processing is performed; Based on the above resolution discrimination processing, pre-processed net load data and pre-processed meteorological data are obtained; The pre-processed net load data and the pre-processed meteorological data are re-indexed based on the time index to ensure that the timestamps of all data sources are aligned, and the net load time series data and the meteorological time series data are obtained accordingly.

4. The power forecasting method taking into account net load power and meteorological factors according to claim 1, characterized in that: The method for obtaining the steady net load time series data is as follows: Constructing a test regression model based on the net load time series data; Estimate the parameter items of the test regression model based on the least squares method to obtain parameter item estimates; Obtaining residual variance based on the estimated values ​​of the parameter terms; Obtaining a corrected standard error based on the residual variance; Obtaining a test statistic based on the parameter estimate and the standard error; determining a critical value, and determining whether the test statistic is less than the critical value; If so, it is determined that the net load time series data has a unit root and is non-stationary data; Otherwise, determining that the net load time series data is stable data and using it as the stable net load time series data; Differential processing is performed based on the non-stationary net load time series data to obtain differential load data, and the above-mentioned stationarity test is cyclically performed based on the differential load data until the stable net load time series data is obtained.

5. The power forecasting method taking into account net load power and meteorological factors according to claim 1, characterized in that: Establishing the optimal SARIMA model specifically includes: Establishing an initial SARIMA model based on the steady net load time series data; Obtaining an optimal model order of the initial SARIMA model based on the steady net load time series data; Substituting the optimal model order into the initial SARIMA model to obtain a general SARIMA model; Perform maximum likelihood estimation based on the general SARIMA model to obtain optimal model parameters; Substituting the optimal model parameters into the general SARIMA model, the optimal SARIMA model is obtained.

6. A power forecasting method taking into account net load power and meteorological factors according to claim 5, characterized in that: The initial SARIMA model is specifically: Non-seasonal part: (1-φ1B-φ2B 2 -...-f p B p (1-B) d Y t =θ0+(θ1B+θ2B 2 +...+θ q B q )e t ; Seasonal section: (1-Φ1B m -Φ2B 2m -...-F P B Pm )(1-B m ) D Y t =Θ0+(Θ1B m +Θ2B 2m +...+Θ Q B Qm )e t ; Among them, φ p represents the coefficient of the pth non-seasonal autoregressive part, p represents the order of non-seasonal autoregressive, B represents the lag term coefficient, d represents the number of non-seasonal differences, Y t represents the value of the steady net load time series data at time t, θ q represents the coefficient of the qth non-seasonal moving average part, q represents the order of the non-seasonal moving average part, ε t represents the error term, Φ p represents the coefficient of the pth non-seasonal autoregressive part, m represents the seasonal period, D represents the number of seasonal differences, Θ Q represents the coefficient of the seasonal moving average part, and Q represents the order of the seasonal moving average part.

7. A power forecasting method taking into account net load power and meteorological factors according to claim 6, characterized in that: The method for obtaining the optimal model order is: Obtaining a seasonal autocorrelation graph, a seasonal partial autocorrelation graph, a non-seasonal autocorrelation graph and a non-seasonal partial autocorrelation graph based on the steady load time series data; Obtaining a candidate order value of a seasonal sliding average and a candidate order value of a seasonal autoregression based on the seasonal autocorrelation graph and the seasonal partial autocorrelation graph; Obtaining corresponding order candidate values ​​of the non-seasonal sliding average and order candidate values ​​of the non-seasonal autoregressive based on the non-seasonal autocorrelation graph and the non-seasonal partial autocorrelation graph; Traversing all the candidate order values ​​based on a step-by-step grid search, fitting the initial SARIMA model to obtain all intermediate SARIMA models; The BLC values ​​are calculated based on all the intermediate SARIMA models, and the model order of the intermediate SARIMA model corresponding to the minimum BLC value is selected as the optimal model order.

8. The power forecasting method taking into account net load power and meteorological factors according to claim 6, characterized in that: The method for obtaining the optimal model parameters is: Based on the general SARIMA model, the initial parameters are obtained using the least squares method; Obtaining residual values ​​of the general SARIMA model based on the initial parameters; constructing a likelihood function for a single observation based on the residual value; Obtaining a joint likelihood function based on all of the likelihood functions; Obtaining a log-likelihood function based on the joint likelihood function; The gradient of the log-likelihood function with respect to each model parameter is calculated based on the gradient descent method, and the model parameters are gradually updated along the gradient descent direction until the log-likelihood function converges to a maximum value, and the finally updated model parameters are used as the optimal model parameters.

9. The power forecasting method taking into account net load power and meteorological factors according to claim 1, characterized in that: Obtain a trained power prediction model, specifically including: Iteratively training the power prediction model based on the training set: The power prediction model calculates a current deviation based on a current predicted value and a true value in each round of iterative training; Training the power prediction model by introducing the current deviation as a new feature into a training set to obtain a new training set; The power prediction model fits the structure of the current deviation adjustment tree in each round of iterative training, and finally obtains the power prediction model that minimizes the loss function as the trained power prediction model.

10. A power forecasting system taking into account net load power and meteorological factors, applied to a power forecasting method taking into account net load power and meteorological factors as claimed in any one of claims 1 to 9, characterized in that: include: Data processing module, stationarity test module, feature extraction module, training set acquisition module, model training module and result output module; The data processing module is used to obtain historical net load data and corresponding meteorological data for a preset period and perform preprocessing to obtain net load time series data and meteorological time series data; The stability test module is used to perform stability test and differential processing based on the net load time series data to obtain stable net load time series data; The feature extraction module is used to establish an optimal SARIMA model based on the steady net load time series data; and extract trend features and seasonal features based on the optimal SARIMA model; The training set acquisition module is used to form a training set based on the meteorological time series data, the trend characteristics and the seasonal characteristics; The model training module is used to construct a power prediction model and perform training based on the training set and the loss function to obtain a trained power prediction model; The result output module is used to obtain future meteorological time series data, historical trend characteristics and historical seasonal characteristics of the target period and input them into the trained power prediction model to obtain the net load power prediction value.

Images