Short-term power load prediction method in combination with weather parameters

Abstract

The invention discloses a short-term power load prediction method in combination with weather parameters. The method comprises the following steps: acquiring power load data and weather parameters of corresponding time; preprocessing the data; using an EEMD method to decompose the power load data into a high-frequency nonlinear component and a low-frequency linear component; in combination with weather parameters, training a CNN-LSTM model which introduces an attention mechanism and is based on multi-scale feature extraction, and outputting a high-frequency nonlinear component predicted value; for the low-frequency linear component, a time sequence model SARIMA is adopted, and a low-frequency linear component predicted value is output; and adding the high-frequency nonlinear component predicted value and the low-frequency linear component predicted value to obtain a complete load prediction result, carrying out error analysis, and carrying out iterative optimization on the model. According to the method, weather data are introduced, high-frequency nonlinear components are extracted through sequence decomposition, weather parameters are used as input to construct a deep learning prediction model, and the prediction capability of complex load fluctuation and the precision and reliability of short-term load prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the field of power load forecasting, relates to short-term power load forecasting technology, and specifically relates to a short-term power load forecasting method combined with weather parameters. Background Art

[0002] With the continuous development of science and technology and social economy, my country's power industry system has also developed tremendously. The continuous development of my country's power system and the increasing modernization and complexity of power grid management have made the research on short-term power system load forecasting more and more attention. The power system has accumulated a large amount of electricity consumption data. Data mining of this massive data can obtain valuable information and provide data support for the diversified services of the power grid. Predicting electricity consumption data is an important application of time series prediction.

[0003] Short-term power load is related to many factors such as weather, economy, day type, and society. Since people will change their electricity consumption behavior as they perceive changes in environmental comfort, such as turning on air conditioning for cooling on hot days and heating on cold days, weather parameters have a significant impact on power load. According to statistics, there is a significant positive correlation between summer temperature and load, and temperature has a cumulative effect on load. Continuous high temperatures for many days will change the characteristics of power load. Therefore, the short-term power load forecasting method combined with weather parameters is worthy of attention. The relationship between weather parameters and power load is a complex nonlinear relationship. Traditional methods have failed to fully explore the nonlinear relationship between weather and load, limiting the accuracy of the forecast. Current load forecasting mostly uses time series analysis, statistical models, or direct modeling based on deep learning. These methods often have shortcomings in dealing with complex nonlinear fluctuations and high-frequency changes in load, and it is difficult to accurately capture the dynamic characteristics of load. Summary of the invention

[0004] Purpose of the invention: In view of the problems that the prior art often has deficiencies in processing complex nonlinear fluctuations and high-frequency changes of loads and it is difficult to accurately capture the dynamic characteristics of loads, a short-term power load forecasting method combined with weather parameters is provided. The high-frequency nonlinear components are extracted by a sequence decomposition method, and a deep learning prediction model is constructed using weather parameters as input, which improves the prediction ability of complex load fluctuations and the accuracy and reliability of short-term load forecasting.

[0005] Technical solution: To achieve the above purpose, the present invention provides a short-term power load forecasting method combined with weather parameters, comprising the following steps:

[0006] S1: Obtain power load data within a preset time period and weather parameters at the corresponding time, including temperature, humidity, rainfall, wind speed, and weather type (sunny, rainy, snowy, etc.);

[0007] S2: Preprocess the data to ensure data quality and consistency;

[0008] S3: Use the EEMD method to decompose the power load data into high-frequency nonlinear components and low-frequency linear components;

[0009] S4: Based on the high-frequency nonlinear component and combined with weather parameters, a CNN-LSTM model based on multi-scale feature extraction with an attention mechanism is trained to output the predicted value of the high-frequency nonlinear component for the time period to be predicted;

[0010] S5: For the low-frequency linear component, the time series model SARIMA is used to capture the long-term trend of the low-frequency linear component and output the predicted value of the low-frequency linear component;

[0011] S6: Add the predicted value of the high-frequency nonlinear component and the predicted value of the low-frequency linear component to obtain a complete load forecast result, perform error analysis between the predicted result and the actual value, and iteratively optimize the model;

[0012] S7: Output the power load prediction result through the optimized model.

[0013] Furthermore, the power load data obtained from the historical data in step S1 obtains historical air temperature, temperature, and rainfall data from the meteorological service platform, and aligns the timestamp with the load data.

[0014] Furthermore, the preprocessing in step S2 includes cleaning, completing, denoising and standardizing the data.

[0015] Furthermore, the method for decomposing the power load data in step S3 is:

[0016] A1: EEMD (Ensembled Empirical Mode Decomposition) improves the robustness and stability of EMD (Empirical Mode Decomposition) by adding random Gaussian white noise.

[0017] Suppose the original signal is X(t), and the signal after adding noise is:

[0018] X i (t)=X(t)+∈ i (t)

[0019] Among them, ∈ i (t) is the i-th superimposed white noise;

[0020] Use EMD to decompose each signal to obtain a set of intrinsic mode functions IMFS. The decomposition results are:

[0021]

[0022] Among them, IMF ij (t) is the jth IMF component of the ith signal, R i (t) is the residual term;

[0023] Superimpose different white noises, repeat the above steps, perform multiple decompositions, average the IMFs of all noise versions, and eliminate the influence of noise on the results; the averaged IMF is:

[0024]

[0025] Where, M is the number of noise superpositions;

[0026] A2: After obtaining the IMF, the zero-crossing rate detection is used to analyze the frequency characteristics of each IMF, especially the high-frequency components. The zero-crossing rate can be calculated by the following formula:

[0027]

[0028] Where N is the number of samples and x[n] is the sample value in the IMF.

[0029] The EEMD method is used to decompose the power load sequence, and the expression is:

[0030] L(t)=L inear (t)+L nonlinear (t)

[0031] Where: L inear (t) is the low-frequency linear component, indicating the long-term trend of the compliance; L nonlinear (t) is the high-frequency nonlinear component, which represents the rapid fluctuation characteristics of the load.

[0032] Furthermore, the modeling and prediction steps of the high-frequency nonlinear component in step S4 are:

[0033] B1: The high-frequency nonlinear component L nonlinear (t) and the weather parameters of the corresponding time period are used as model inputs; the weather parameters are organized into a matrix by time step:

[0034] X weather ={X weather,t ,X weather,t+1 ,…,X weather,T}

[0035] Each X weather,t is a weather parameter vector;

[0036] The high-frequency nonlinear component is the high-frequency component extracted from the power load time series by EEMD:

[0037] X high ={Xhigh,t ,X high,t+1 ,…,X high,T}

[0038] Then the weather parameters and high-frequency components are normalized to [0,1];

[0039] B2: The CNN-LSTM model based on multi-scale feature extraction is composed of three modules: CNN, LSTM and Attention mechanism. According to the weather parameters input in step B1, CNN extracts the local features of the weather parameters and identifies the correlation between weather and load fluctuations. According to the high-frequency nonlinear components input in step B1, LSTM captures the time series characteristics of the high-frequency components and models their dynamic changes. The attention mechanism is used to improve the model's key capabilities for key time steps and important weather parameters.

[0040] B3: Use historical power load data to construct training set data and validation set data, use mean square error as the loss function, dynamically adjust the learning rate through the Adam optimization algorithm to improve training efficiency, and obtain a trained CNN-LSTM model.

[0041] Furthermore, the CNN network model structure in the CNN-LSTM model of step B2 includes a convolution layer and a pooling layer. The convolution layer extracts the spatial local pattern of weather parameters, such as the coupling relationship between temperature and humidity; the pooling layer is used to reduce the feature dimension while retaining the key pattern; CNN extracts the local feature F of weather parameters. weather , which can reflect the short-term impact of temperature changes on load, the interaction of parameters such as wind speed and humidity;

[0042] F weather ={f1,f2,…,fd}

[0043] Among them, f i is the i-th weather feature;

[0044] score(f i )=W weather *f i +b weather

[0045] Among them, W weather and b weather is a learning parameter.

[0046] Furthermore, the LSTM network model in the CNN-LSTM model of step B2 models the high-frequency load component to capture the dynamic changes in time. The LSTM network model structure includes a forget gate, an input gate, and an output gate. The feature H extracted by LSTM high Represents the hidden state sequence of the generated time steps, including periodic fluctuations and random characteristics or abnormal point changes;

[0047] H high ={h1,h2,…,hT}

[0048] score(h t )=O t tanh(C t )

[0049] Among them, O t is the output gate result, C t It is the result of the unit updating status.

[0050] Furthermore, the operation of the attention mechanism in the CNN-LSTM model of step B2 includes:

[0051] Calculate the importance of features using a fully connected network:

[0052]

[0053] Generate attention-weighted weather features:

[0054]

[0055] Calculate the attention score through context information and hidden state:

[0056]

[0057] Generating attention-weighted time series features

[0058]

[0059] After integrating the features of the attention mechanism, feature splicing is performed:

[0060] F combined =Concat(F weather,attn ,H high,attn )

[0061] Fully connected output:

[0062] y t =W out *F combined +b out .

[0063] The purpose of introducing the attention mechanism is to assign weights to different time steps or feature components so that the model can focus more on important input information. Dynamic weights are assigned to local features of weather parameters to focus on important meteorological factors, such as sudden changes in temperature or humidity. The time steps of high-frequency load components are weighted to highlight the fluctuation characteristics of key time periods.

[0064] Furthermore, in step S5, for the low-frequency linear component L linear (t) Independent forecasting using the SARIMA model.

[0065] Furthermore, the error analysis in step S6 is crucial, which is not only a key step in evaluating the performance of the model, but also an important basis for discovering the potential change direction of the model. The error analysis method includes:

[0066] D1: Calculate the difference between the actual load value and the predicted load value;

[0067] D2: Common evaluation indicators include mean square error (MSE), mean absolute error (MAE), absolute percentage error (MAPE), and absolute coefficient R, which measures the degree of fit between the predicted load value and the actual load value. 2 ;

[0068] D3: Use visualization tools to draw time series graphs of actual load values, predicted load values, and errors to observe the fluctuation trend of errors; use box plots to display the distribution characteristics of errors and find the concentrated intervals of abnormal values.

[0069] Beneficial effects: Compared with the prior art, the present invention combines EEMD sequence decomposition and zero-crossing rate detection methods to decompose signals and analyze the frequency characteristics of each IMF, which can deeply understand the changes in signals. The extracted high-frequency nonlinear components and weather parameters are used as input to build a deep learning combined prediction model. The attention mechanism is added to the CNN-LSTM model to process the relationship between high-frequency nonlinear components and weather parameters, which improves the prediction ability of complex load fluctuations and the accuracy and reliability of short-term load forecasting.

[0070] Attached photos

[0071] Figure 1 It is a schematic diagram of the process of the present invention;

[0072] Figure 2 It is a schematic diagram of the process of sequence decomposition in the present invention;

[0073] Figure 3 It is a structural diagram of the CNN-LSTM model in the present invention. DETAILED DESCRIPTION

[0074] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0075] The present invention provides a short-term power load forecasting method combining weather parameters, such as Figure 1 As shown, it includes the following steps:

[0076] S1: Obtain power load data within a preset time period and weather parameters at the corresponding time, including temperature, humidity, rainfall, wind speed, and weather type (sunny, rainy, snowy, etc.);

[0077] The power load data obtained from historical data has a time granularity of hours. The historical air temperature, temperature, rainfall and other data are obtained from the meteorological service platform and the timestamps are aligned with the load data.

[0078] S2: Clean, complete, denoise and standardize the data to ensure data quality and consistency;

[0079] S3: Use the EEMD method to decompose the power load data into high-frequency nonlinear components and low-frequency linear components;

[0080] like Figure 2 As shown in Figure 2, the method for decomposing power load data is:

[0081] A1: EEMD (Ensembled Empirical Mode Decomposition) improves the robustness and stability of EMD (Empirical Mode Decomposition) by adding random Gaussian white noise.

[0082] Suppose the original signal is X(t), and the signal after adding noise is:

[0083] X i (t)=X(t)+∈ i (t)

[0084] Among them, ∈ i (t) is the i-th superimposed white noise;

[0085] Use EMD to decompose each signal to obtain a set of intrinsic mode functions IMFS. The decomposition results are:

[0086]

[0087] Among them, IMF ij (t) is the jth IMF component of the ith signal, R i (t) is the residual term;

[0088] Superimpose different white noises, repeat the above steps, perform multiple decompositions, average the IMFs of all noise versions, and eliminate the influence of noise on the results; the averaged IMF is:

[0089]

[0090] Where, M is the number of noise superpositions;

[0091] A2: After obtaining the IMF, the zero-crossing rate detection is used to analyze the frequency characteristics of each IMF, especially the high-frequency components. The zero-crossing rate is calculated by the following formula:

[0092]

[0093] Where N is the number of samples and x[n] is the sample value in the IMF.

[0094] High-frequency IMFs usually show a higher zero-crossing rate, and low-frequency IMFs usually show a lower zero-crossing rate. The power load sequence can be decomposed to obtain the expression:

[0095] L(t)=L linear (t)+L nonlinear (t)

[0096] Where: L linear (t) is the low-frequency linear component, indicating the long-term trend of the compliance; L nonlinear (t) is the high-frequency nonlinear component, which represents the rapid fluctuation characteristics of the load.

[0097] S4: Based on the high-frequency nonlinear component and combined with weather parameters, a CNN-LSTM model based on multi-scale feature extraction with an attention mechanism is trained to output the predicted value of the high-frequency nonlinear component for the time period to be predicted;

[0098] The steps for modeling and predicting high-frequency nonlinear components are:

[0099] B1: The high-frequency nonlinear component L nonlinear (t) and the weather parameters of the corresponding time period are used as model inputs; the weather parameters are organized into a matrix by time step:

[0100] X weather ={X weather,t ,X weather,t+1 ,…,X weather,T}

[0101] Each X weather,t is a weather parameter vector;

[0102] The high-frequency nonlinear component is the high-frequency component extracted from the power load time series by EEMD:

[0103] X high ={X high,t ,X high,t+1 ,…,X high,T}

[0104] Then the weather parameters and high-frequency components are normalized to [0,1];

[0105] B2: If Figure 3 As shown in the figure, the CNN-LSTM model based on multi-scale feature extraction is composed of three modules: CNN, LSTM and Attention mechanism. According to the weather parameters input in step B1, CNN extracts the local features of the weather parameters and identifies the correlation between weather and load fluctuations. According to the high-frequency nonlinear components input in step B1, LSTM captures the time series characteristics of the high-frequency components and models their dynamic changes. The attention mechanism is used to improve the model's key capabilities for key time steps and important weather parameters.

[0106] The CNN network model structure in the CNN-LSTM model includes convolutional layers and pooling layers. The convolutional layers extract the spatial local patterns of weather parameters, such as the coupling relationship between temperature and humidity; the pooling layer is used to reduce the feature dimension while retaining the key patterns; CNN extracts the local features F of weather parameters. weather , which can reflect the short-term impact of temperature changes on load, the interaction of parameters such as wind speed and humidity;

[0107] F wetather ={f1,f2,…,fd}

[0108] Among them, f i is the i-th weather feature;

[0109] score(f i )=W weather *f i +b weather

[0110] Among them, W weather and b weather is a learning parameter.

[0111] The LSTM network model in the CNN-LSTM model models high-frequency load components and captures dynamic changes in time. The LSTM network model structure includes a forget gate, an input gate, and an output gate. The feature H extracted by LSTM high Represents the hidden state sequence of the generated time steps, including periodic fluctuations and random characteristics or abnormal point changes;

[0112] H high ={h1,h2,…,hT}

[0113] score(h t )=O t tanh(C t )

[0114] Among them, O t is the output gate result, C t It is the result of the unit updating status.

[0115] The operation of the attention mechanism in the CNN-LSTM model includes:

[0116] Calculate the importance of features using a fully connected network:

[0117]

[0118] Generate attention-weighted weather features:

[0119]

[0120] Calculate the attention score through context information and hidden state:

[0121]

[0122] Generating attention-weighted time series features

[0123]

[0124] After integrating the features of the attention mechanism, perform feature concatenation:

[0125] F combined =Concat(F weather ,attn,H high ,attn)

[0126] Fully connected output:

[0127] y t =W out *F combined +b out .

[0128] The purpose of introducing the attention mechanism is to assign weights to different time steps or feature components so that the model can focus more on important input information. Dynamic weights are assigned to local features of weather parameters to focus on important meteorological factors, such as sudden changes in temperature or humidity. The time steps of high-frequency load components are weighted to highlight the fluctuation characteristics of key time periods.

[0129] B3: Use historical power load data to construct training set data and validation set data, use mean square error as the loss function, dynamically adjust the learning rate through the Adam optimization algorithm to improve training efficiency, and obtain a trained CNN-LSTM model.

[0130] S5: for low-frequency linear component L linear (t) Using the SARIMA seasonal model for independent prediction, capturing the long-term trend of the low-frequency linear component, and outputting the predicted value of the low-frequency linear component;

[0131] S6: Construct a composite model composed of linear and nonlinear components, add the predicted values ​​of high-frequency nonlinear components to the predicted values ​​of low-frequency linear components to obtain a complete load forecast result, perform error analysis between the predicted results and the actual values, and iteratively optimize the model. This makes up for the problem that a single model cannot predict the complexity of the power load sequence;

[0132] Error analysis is crucial. It is not only a key step in evaluating model performance, but also an important basis for discovering potential changes in the model. Error analysis methods include:

[0133] D1: Calculate the difference between the actual load value and the predicted load value;

[0134] D2: Common evaluation indicators include mean square error (MSE), mean absolute error (MAE), absolute percentage error (MAPE), and absolute coefficient R, which measures the degree of fit between the predicted load value and the actual load value. 2 ;

[0135] D3: Use visualization tools to draw time series graphs of actual load values, predicted load values, and errors to observe the fluctuation trend of errors; use box plots to display the distribution characteristics of errors and find the concentrated intervals of abnormal values.

[0136] S7: Output the power load prediction result through the optimized model.

[0137] According to the above content, short-term power load forecasting is a key task in the operation of the power system and is widely used in fields such as power generation planning, load scheduling and power market transactions. The fluctuation of power load is affected by many factors, including the inherent laws of historical load (linear and nonlinear characteristics) and external weather conditions (temperature, humidity, etc.). The present invention adopts the overall framework of "sequence decomposition + feature extraction + deep learning prediction" and proposes a high-precision and robust short-term power load forecasting method by combining high-frequency nonlinear components with weather parameters. This method can not only capture the complex fluctuation characteristics of the load, but also fully explore the driving role of weather parameters, and has significant theoretical and engineering application value.

[0138] This embodiment also provides a short-term power load forecasting system combined with weather parameters, which includes a network interface, a memory and a processor; wherein the network interface is used to realize the reception and transmission of signals in the process of sending and receiving information between other external network elements; the memory is used to store computer program instructions that can be run on the processor; the processor is used to execute the steps of the above-mentioned consensus method when running the computer program instructions.

[0139] The present embodiment also provides a computer storage medium, which stores a computer program, and the method described above can be implemented when the processor executes the computer program. The computer readable medium can be considered to be tangible and non-temporary. Non-limiting examples of non-temporary tangible computer-readable media include non-volatile memory circuits (such as flash memory circuits, erasable programmable read-only memory circuits or mask read-only memory circuits), volatile memory circuits (such as static random access memory circuits or dynamic random access memory circuits), magnetic storage media (such as analog or digital tapes or hard drives) and optical storage media (such as CDs, DVDs or Blu-ray discs), etc. The computer program includes processor executable instructions stored on at least one non-temporary tangible computer-readable medium. The computer program may also include or rely on stored data. The computer program may include a basic input / output system (BIOS) that interacts with the hardware of a special-purpose computer, a device driver that interacts with a specific device of a special-purpose computer, one or more operating systems, user applications, background services, background applications, etc.

[0140] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

Claims

1. A short-term power load forecasting method combining weather parameters, characterized in that: The steps include: S1: Obtaining power load data within a preset time period and weather parameters at the corresponding time; S2: Preprocess the data to ensure data quality and consistency; S3: Use the EEMD method to decompose the power load data into high-frequency nonlinear components and low-frequency linear components; S4: Based on the high-frequency nonlinear component and combined with weather parameters, a CNN-LSTM model based on multi-scale feature extraction with an attention mechanism is trained to output the predicted value of the high-frequency nonlinear component for the time period to be predicted; S5: For the low-frequency linear component, the time series model SARIMA is used to capture the long-term trend of the low-frequency linear component and output the predicted value of the low-frequency linear component; S6: Add the predicted value of the high-frequency nonlinear component and the predicted value of the low-frequency linear component to obtain a complete load forecast result, perform error analysis between the predicted result and the actual value, and iteratively optimize the model; S7: Output the power load prediction result through the optimized model.

2. A short-term power load forecasting method combining weather parameters according to claim 1, characterized in that: The power load data obtained from the historical data in step S1 obtains the historical air temperature, temperature, and rainfall data from the meteorological service platform, and aligns the timestamp with the load data.

3. A short-term power load forecasting method combining weather parameters according to claim 1, characterized in that: The preprocessing in step S2 includes cleaning, completing, denoising and standardizing the data.

4. A short-term power load forecasting method combining weather parameters according to claim 1, characterized in that: The method for decomposing the power load data in step S3 is: A1: Let the original signal be X(t), and the signal after adding noise is: X i (t)=X(t)+∈ i (t) Among them, ∈ i (t) is the i-th superimposed white noise; Use EMD to decompose each signal to obtain a set of intrinsic mode functions IMFS. The decomposition results are: Among them, IMF ij (t) is the jth IMF component of the ith signal, R i (t) is the residual term; Superimpose different white noises, repeat the above steps, perform multiple decompositions, average the IMFs of all noise versions, and eliminate the influence of noise on the results; the averaged IMF is: Where, M is the number of noise superpositions; A2: After obtaining the IMF, the frequency characteristics of each IMF are analyzed using zero-crossing rate detection. The zero-crossing rate is calculated using the following formula: Where N is the number of samples, x[n] is the sample value in IMF; The EEMD method is used to decompose the power load sequence, and the expression is: L(t)=L linear (t)+L nonlinear (t) Where: L linear (t) is the low-frequency linear component, indicating the long-term trend of the compliance; L nonlinear (t) is the high-frequency nonlinear component, which represents the rapid fluctuation characteristics of the load.

5. A short-term power load forecasting method combining weather parameters according to claim 4, characterized in that: The modeling and prediction steps of the high-frequency nonlinear component in step S4 are: B1: The high-frequency nonlinear component L nonlinear (t) and the weather parameters of the corresponding time period are used as model inputs; the weather parameters are organized into a matrix by time step: X weather ={X weather,t ,X weather,t +1,...,X weather,T } Each X weather,t is a weather parameter vector; The high-frequency nonlinear component is the high-frequency component extracted from the power load time series by EEMD: X high ={X high,t ,X high,t+1 ,...,X high,T } Then the weather parameters and high-frequency components are normalized to [0, 1]; B2: The CNN-LSTM model based on multi-scale feature extraction is composed of three modules: CNN, LSTM and Attention mechanism. According to the weather parameters input in step B1, CNN extracts the local features of the weather parameters and identifies the correlation between weather and load fluctuations. According to the high-frequency nonlinear components input in step B1, LSTM captures the time series characteristics of the high-frequency components and models their dynamic changes. The attention mechanism is used to improve the model’s critical capabilities for key time steps and important weather parameters; B3: Use historical power load data to construct training set data and validation set data, use mean square error as the loss function, dynamically adjust the learning rate through the Adam optimization algorithm to improve training efficiency, and obtain a trained CNN-LSTM model.

6. A short-term power load forecasting method combining weather parameters according to claim 5, characterized in that: The CNN network model structure in the CNN-LSTM model of step B2 includes a convolution layer and a pooling layer. The convolution layer extracts the spatial local pattern of the weather parameters, and the pooling layer is used to reduce the feature dimension while retaining the key pattern. CNN extracts the local feature F of the weather parameters. weather : F weather ={f1,f2,...,fd} Among them, f i is the i-th weather feature; score(f i )=W weather *f i +b weather Among them, W weather and b weather is a learning parameter.

7. A short-term power load forecasting method combining weather parameters according to claim 5, characterized in that: The LSTM network model in the CNN-LSTM model of step B2 models the high-frequency load component and captures the dynamic changes in time. The LSTM network model structure includes a forget gate, an input gate, and an output gate. The feature H extracted by LSTM high Represents the hidden state sequence of the generated time steps, including periodic fluctuations and random characteristics or abnormal point changes; H high ={h1,h2,…,hT} score(h t )=O t fishy(C) t ) Among them, O t is the output gate result, C t It is the result of the unit updating status.

8. A short-term power load forecasting method combining weather parameters according to claim 5, characterized in that: The operation of the attention mechanism in the CNN-LSTM model of step B2 includes: Calculate the importance of features using a fully connected network: Calculate the attention score through context information and hidden state: Generating attention-weighted time series features After integrating the features of the attention mechanism, feature splicing is performed: F combined =Concat(F weather,attn ,H high,attn ) Fully connected output: y t =W out *F combined +b out 。 9. A method and system for short-term power load forecasting combined with weather parameters according to claim 1, characterized in that: In step S5, the low-frequency linear component L linear (t) Independent forecasting using the SARIMA model.

10. A method and system for short-term power load forecasting combined with weather parameters according to claim 1, characterized in that: The error analysis method in step S6 includes: D1: Calculate the difference between the actual load value and the predicted load value; D2: The evaluation indicators include mean square error, mean absolute error, absolute percentage error and absolute coefficient R to measure the degree of fit between the predicted load value and the actual load value. 2 ; D3: Use visualization tools to draw time series graphs of actual load values, predicted load values, and errors, observe the fluctuation trend of errors, display the distribution characteristics of errors through box plots, and find the concentrated intervals of abnormal values.

Images