Multi-source information integration runoff prediction method based on informer model

Abstract

The invention discloses a runoff prediction model based on multi-source information integration of an Informer model, and belongs to the technical field of hydropower station runoff prediction. The invention aims to solve the problem of poor accuracy of medium and long term runoff prediction, especially for runoff sequences with multi-source, non-linear and non-stationary characteristics. The monitoring power station range is selected, and historical runoff data of the hydropower station and hydrological information such as rainfall and temperature of a surrounding area are obtained; adopting a Pearson correlation analysis method to select strong correlation features as multi-source time sequence feature input of an Informer model; preprocessing the data by using an STL seasonal decomposition method, and extracting characteristics of a period term and a trend term; runoff prediction is carried out by adopting a sparse attention mechanism and a self-attention distillation mechanism of an Informer model; according to the method, the prediction speed and feature capture can be enhanced, the runoff prediction accuracy is improved, and powerful support is provided for water resource management and scheduling of the hydropower station.

Description

Technical Field

[0001] The present invention relates to the technical field of runoff prediction, and in particular to a runoff prediction method based on multi-source information integration of an informer model. Background Art

[0002] Runoff prediction is an important technical link in the fields of hydropower station operation management, water resources scheduling, and flood prevention and disaster reduction. Accurate runoff prediction can help decision makers formulate reasonable water resource utilization plans, optimize the power generation efficiency of hydropower stations, and reduce the occurrence of flood disasters, which is of great significance for ensuring the sustainable development of social economy.

[0003] Among the existing runoff prediction technologies, traditional methods mainly include machine learning methods such as support vector machines and random forests. These methods can improve the accuracy of runoff prediction to a certain extent. However, when facing medium- and long-term runoff prediction, due to the large number of influencing factors and the strong nonlinearity and non-stationarity of the runoff process, traditional methods often find it difficult to accurately capture the complex laws of runoff changes, resulting in unsatisfactory prediction results.

[0004] In recent years, with the rapid development of artificial intelligence technology, prediction methods based on deep learning have gradually been applied to the field of runoff prediction. Through multiple layers of nonlinear transformation, deep learning models can automatically learn complex features in data, showing certain advantages in medium- and long-term runoff prediction. However, existing deep learning models still face challenges when processing long sequence data, such as low computational efficiency and insufficient feature capture capabilities, which limits their widespread application in the field of runoff prediction.

[0005] In addition, although some studies have attempted to introduce multi-source information (such as rainfall, temperature, etc.) into the runoff prediction model to improve the prediction accuracy, for example, in the article "Multi-scale wavelet analysis of precipitation and runoff series in the Yellow River source area during the flood season" in the 4th issue of "People's Yellow River" in 2024, wavelet analysis was used to perform multi-time scale analysis on single runoff data, and the periodic law, change trend and correlation of runoff were obtained; however, due to the access of a large number of data acquisition devices, the explosive growth of multi-dimensional data and the increasing requirements for prediction accuracy, the classic parameter model and traditional machine learning algorithm can no longer meet the high efficiency and high precision requirements of the prediction task; and these methods often lack an effective information integration mechanism, making it difficult to fully explore and utilize the correlation and complementarity between multi-source information. Therefore, how to construct a runoff prediction model that can efficiently process long sequence data and effectively integrate multi-source information has become an important issue that needs to be solved in the current field of runoff prediction technology.

[0006] In summary, the existing technology still has many deficiencies in medium- and long-term runoff prediction, and a more efficient and accurate prediction method is needed to solve these problems. The present invention is proposed in this context, aiming to improve the accuracy and stability of medium- and long-term runoff prediction by introducing the Informer model and multi-source information integration technology. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a runoff prediction method based on multi-source information integration of an informer model, to solve the problem of poor accuracy of medium- and long-term runoff prediction in the field of runoff prediction technology, especially for the prediction of runoff sequences with multi-source, nonlinear and non-stationary characteristics, to overcome the limitation of the prior art in the poor prediction results of original long sequence data.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is a runoff prediction method based on multi-source information integration of an informer model, comprising the following steps: Step 1: Select the monitoring power station range and obtain the historical runoff data of the hydropower station and the hydrological information such as rainfall and temperature in the surrounding area; Step 2: Using the Pearson correlation analysis method, the features that are strongly correlated with the historical runoff data of the hydropower station are selected from the acquired hydrological information as the multi-source time series feature input of the informer model; Step 3: Use the STL (Seasonal-Trend Decomposition using Loess) seasonal decomposition method to preprocess the selected data set and extract the characteristics of the periodic term and trend term; Step 4: Use the Informer model to predict runoff on the preprocessed data using the sparse attention mechanism and self-attention distillation mechanism to obtain the predicted runoff value at the future moment.

[0009] In a preferred solution, in Step 1, the selected monitoring power station range includes at least one hydropower station and rainfall stations and hydrological stations in its upstream basin to improve prediction accuracy.

[0010] In the preferred solution, in Step 2, the specific implementation of the Pearson correlation analysis method includes calculating the Pearson correlation coefficient between each feature and the target prediction value, and selecting features whose absolute values ​​are greater than a preset threshold as strong correlation features, and the preset threshold is 0.4.

[0011] In the preferred embodiment, the calculation formula of the Pearson correlation coefficient is as follows: (1); In the formula, is the i-th data of the j-th data to be predicted, is the jth need and historical data The i-th data of the hydrological information to be analyzed, and They are and The mean of yes and The correlation coefficient of .

[0012] In a preferred solution, in Step 3, the STL seasonal decomposition method decomposes the original data into trend terms, seasonal terms and residual terms, wherein the decomposition period T is set to 365 days to ensure accurate extraction of seasonal characteristics.

[0013] In a preferred embodiment, the detailed steps of the STL seasonal decomposition method include: Step 3.1: Divide the original data into multiple subsequences according to the size of the decomposition period, and then perform local polynomial smoothing on the subsequences to remove noise and rapidly changing parts to obtain a preliminary trend estimate; Step 3.2: Calculate the seasonal average at each time to estimate the seasonal component, remove the trend and seasonal data by subtraction, and perform smoothing again to obtain the residual component; Step 3.3: Repeat the decomposition several times until the estimated values ​​of the trend term and seasonal term converge.

[0014] In the preferred solution, in Step 4, the Informer model models the input multi-source time series features through the encoder, generates runoff prediction results using the decoder, and improves the prediction speed and accuracy through the self-attention distillation mechanism.

[0015] In a preferred embodiment, the specific steps of the Informer model modeling the input multi-source time series features through the encoder include: using the encoder in the model to model the input time series data and additional seasonal and trend features, and then using the probabilistic self-attention mechanism and the global-local information fusion mechanism to effectively capture the global dependencies and local patterns of the input sequence; based on the output of the encoder, using the decoder to generate the runoff prediction result.

[0016] The preferred solution also includes adjusting the water resource management and scheduling strategy of the hydropower station according to the prediction results to optimize the power generation plan and water storage management.

[0017] The runoff prediction method based on multi-source information integration of the informer model provided by the present invention has the following beneficial effects: 1. The present invention solves the problem of poor accuracy of medium- and long-term runoff prediction in the field of runoff prediction technology, especially for the prediction of runoff series with multi-source, nonlinear and non-stationary characteristics, and overcomes the limitation of poor prediction results of original long sequence data in the prior art; 2. The present invention improves the accuracy of prediction by integrating historical runoff data of the hydropower station and various hydrological information such as rainfall and temperature in the surrounding area; 3. The present invention uses the Pearson correlation analysis method to remove the features with weak correlation with the target prediction value, and uses the STL seasonal decomposition method to pre-process the data to extract the periodic term and trend term features, further enhancing the prediction ability of the model; 4. The present invention adopts the sparse attention mechanism and self-attention distillation mechanism of the Informer model for runoff prediction, effectively capturing the global dependencies and local patterns of the input sequence, and improving the prediction speed and feature capture capability; 5. The present invention adopts the Informer model for prediction, which effectively captures the global dependencies and local patterns of the input sequence, and enhances the robustness and generalization ability of the model; 6. The accurate runoff prediction of the present invention provides a scientific basis and decision-making support for the operation and management of the hydropower station, which helps to optimize water resource allocation and improve power generation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below in conjunction with the accompanying drawings and implementation examples: Figure 1 Schematic diagram of the training process of the prediction model of the present invention; Figure 2 It is the Pearson correlation analysis diagram of the hydrological information of the present invention; Figure 3 It is the STL seasonal decomposition flow chart of the present invention; Figure 4 This is a seasonal decomposition diagram of the historical runoff data of the hydropower station to be measured in Example 2 of the present invention. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments: Example 1 like Figures 1 to 3 As shown in FIG. 1 , the runoff prediction method based on multi-source information integration of the informer model includes the following steps: Step 1: Select the monitoring power station range and obtain the historical runoff data of the hydropower station and the hydrological information such as rainfall and temperature in the surrounding area; Step 2: Using the Pearson correlation analysis method, the features that are strongly correlated with the historical runoff data of the hydropower station are selected from the acquired hydrological information as the multi-source time series feature input of the informer model; Step 3: Use the STL seasonal decomposition method to preprocess the selected data set and extract the characteristics of periodic items and trend items; Step 4: Use the Informer model to predict the runoff of the preprocessed data using the sparse attention mechanism and self-attention distillation mechanism to obtain the predicted runoff value at the future moment.

[0020] In this embodiment, in Step 1, the selected monitoring power station range includes at least one hydropower station and a rainfall station and a hydrological station in its upstream basin to improve the prediction accuracy.

[0021] Furthermore, in Step 2, the specific implementation of the Pearson correlation analysis method includes calculating the Pearson correlation coefficient between each feature and the target prediction value, and selecting features whose absolute values ​​are greater than a preset threshold as strong correlation features, and the preset threshold is 0.4.

[0022] Furthermore, the calculation formula of the Pearson correlation coefficient is as follows: (1) In the formula, is the i-th data of the j-th data to be predicted, is the jth need and historical data The i-th data of the hydrological information to be analyzed, and They are and The mean of yes and The correlation coefficient of .

[0023] Furthermore, in Step 3, the STL seasonal decomposition method decomposes the original data into trend terms, seasonal terms and residual terms, wherein the decomposition period T is set to 365 days to ensure accurate extraction of seasonal features.

[0024] Furthermore, the detailed steps of the STL seasonal decomposition method include: Step 3.1: Divide the original data into multiple subsequences according to the size of the decomposition period, and then perform local polynomial smoothing on the subsequences to remove noise and rapidly changing parts to obtain a preliminary trend estimate; Step 3.2: Calculate the seasonal average at each time to estimate the seasonal component, remove the trend and seasonal data by subtraction, and perform smoothing again to obtain the residual component; Step 3.3: Repeat the decomposition several times until the estimated values ​​of the trend term and seasonal term converge.

[0025] Furthermore, in the Step 4, the Informer model models the input multi-source time series features through the encoder, generates runoff prediction results using the decoder, and improves the prediction speed and accuracy through the self-attention distillation mechanism.

[0026] Furthermore, the specific steps of the Informer model modeling the input multi-source time series features through the encoder include: using the encoder in the model to model the input time series data and additional seasonal and trend features, and then using the probabilistic self-attention mechanism and the global-local information fusion mechanism to effectively capture the global dependencies and local patterns of the input sequence; based on the output of the encoder, using the decoder to generate the runoff prediction result.

[0027] Furthermore, it also includes adjusting the water resource management and scheduling strategies of the hydropower station according to the prediction results to optimize the power generation plan and water storage management.

[0028] Example 2 In another preferred embodiment, based on the above embodiment 1, Figures 1 to 3 As shown, this embodiment describes in detail the implementation steps of the runoff prediction method based on multi-source information integration of the informer model, which specifically includes: S1. Data acquisition and preprocessing: Select the monitoring power station range, for example, hydropower station A, and obtain the historical runoff data of the hydropower station, as well as hydrological information such as rainfall and temperature in the surrounding areas; in order to improve the accuracy of the prediction, select the historical runoff data of hydropower station A and the historical runoff data and rainfall of another power station B located upstream, the temperature of the area where hydropower station A is located, the rainfall data of the main water inflow area, and the precipitation of the five surrounding areas; these data are collected by professional hydrological monitoring stations, and the inverse distance weighted method is used to interpolate the national daily rainfall raster data, and then the daily precipitation raster data within the boundaries of prefecture-level cities in my country is averaged to obtain the daily rainfall in prefecture-level cities.

[0029] S2. Feature selection and correlation analysis: The Pearson correlation analysis method is used to select features of the acquired hydrological information; Pearson correlation analysis plays a key role in the feature selection process, and can remove those features that are weakly correlated with the target prediction value (i.e., the runoff of hydropower station A); the value of the correlation coefficient is between -1 and +1. If the absolute value of the Pearson coefficient between two features is less than 0.4, it is considered that the correlation between the two features is weak and should be excluded. In this way, it can be ensured that only features that are highly correlated with the target prediction value are used in the subsequent prediction model.

[0030] S3, STL seasonal decomposition method preprocessing: The selected data set is preprocessed using the STL seasonal decomposition method; the STL seasonal decomposition method is used to identify and analyze different components in the time series, including trend terms, seasonal terms, and residual terms. The specific steps are as follows: S3.1. Initial decomposition: Set the decomposition period T to 365 days, smooth the time series data, and obtain preliminary estimates of trend and seasonal components; S3.2, Seasonal smoothing: Subtract the initial trend from the time series to obtain the detrended series, smooth each seasonal cycle position separately, and use the Loess (Locally Weighted Scatterplot Smoothing) method to estimate the mean value of each seasonal position to form a preliminary seasonal component; S3.3, Trend smoothing: Subtract the initial seasonal component from the original sequence to obtain the deseasonalized sequence, perform Loess smoothing on the deseasonalized sequence, and estimate the new trend component; S3.4. Residual calculation: Subtract the newly estimated trend and seasonal components from the original time series to obtain the residual; S3.5, Iteration: Repeat the above steps until the estimates of trend and seasonal components converge or the specified number of iterations is reached.

[0031] Taking the data of a power station as an example, the components of the time series after decomposition are analyzed by STL. Figure 4 As shown, we can see that the trend term, seasonal term, and residual term are clearly separated.

[0032] S4, Informer model prediction: The data preprocessed by the STL seasonal decomposition method and the related seasonal and trend characteristics are used as the input of the Informer model; the Informer model uses an encoder to model the input time series data, and then uses the probabilistic self-attention mechanism and self-attention distillation mechanism to predict runoff, which specifically includes the following parts: S4.1, Probabilistic sparse self-attention mechanism: The KL (Kullback-Leibler Divergence) divergence measures the distribution distance and determines the sparsity of the query vector, thereby optimizing the self-attention calculation process and improving the calculation efficiency; S4.2, Self-attention distillation mechanism: Use distillation operations at each layer to extract dominant features and generate compact self-attention feature maps in the next layer, further reducing the amount of computation and maintaining prediction performance; S4.3. Generative decoder structure: A standard decoding structure is adopted, consisting of two identical multi-head attention layers and a feedforward network to produce prediction output.

[0033] According to the output of the encoder, the decoder is used to generate the runoff prediction results of hydropower station A at future moments.

[0034] In order to verify the effectiveness of STL's decomposition of runoff signals for time series prediction models, the following two processing methods were used to compare the original data training data: The runoff signal is not processed, and the Informer model is used directly for runoff prediction.

[0035] The STL seasonal decomposition method is used to process the runoff data, and then the Informer model is used to predict the runoff.

[0036] The experimental results show that the use of STL has greatly improved the Informer prediction accuracy, which proves that the introduction of STL into the Informer model can better capture and predict the changes in runoff, thereby improving the accuracy of runoff prediction.

[0037] Example 3 In another preferred embodiment, based on the above embodiment 1, Figures 1 to 3 As shown, this embodiment further uses an example to specifically illustrate the implementation steps of the runoff prediction method based on multi-source information integration of the informer model, including: S1. Obtain the original data from each hydrological station, use the inverse distance weighted method to interpolate the daily precipitation grid data across the country, and average the daily precipitation grids within the boundaries of prefecture-level cities in my country to obtain the daily precipitation in prefecture-level cities.

[0038] S2. In this example, the measured daily runoff data set from June 1, 2006 to June 1, 2022 at the hydrological station (a total of 5845 sets, with a collection interval of 1 day) is used. The data set is divided into training set, validation set and test set in a ratio of 8:1:1; before training, the min-max method is used to normalize the original data, and the calculation formula is as follows: (2) In the formula, is the hydrological data before scaling, is the minimum hydrological data collected, is the largest hydrological data collected, is the scaled hydrological data.

[0039] S3,Pearson correlation analysis can remove those features that are weakly correlated with the target prediction value; Figure 2 As shown in the figure, if the absolute value of the Pearson coefficient between two features is less than 0.4, it is considered that the correlation between the two features is weak, and the weakly correlated data will be removed. The calculation formula of Pearson is as follows: (1) S4, STL seasonal decomposition method, used to identify and analyze different components in time series, help us understand the basic structure and rules of data, and can handle time series data with noise or irregularity, suitable for seasonal and long-term trend analysis. Its flowchart is as follows Figure 3 As shown, the specific steps include: S4.1. Initial decomposition: The decomposition period T is set to 365. First, the time series data are smoothed to obtain preliminary estimates of the trend and seasonal components; S4.2, Seasonal smoothing: Subtract the initial trend from the time series to obtain the detrended series. Smooth the position of each seasonal cycle separately, and use Loess to estimate the mean value of each seasonal position to form a preliminary seasonal component; S4.3, Trend smoothing: Subtract the initial seasonal component from the original sequence to obtain the deseasonalized sequence, perform Loess smoothing on the deseasonalized sequence, and estimate the new trend component; S4.4. Residual calculation: Subtract the newly estimated trend and seasonal components from the original time series to obtain the residual; S4.5, Iteration: Repeat the above steps until the estimates of trend and seasonal components converge or the specified number of iterations is reached.

[0040] Taking the data of a power station as an example, the components of the time series after decomposition are analyzed by STL. Figure 4 shown.

[0041] S5. Import the processed data into the informer model for input prediction. In this model, the encoder is used to model the input time series data and the additional seasonal and trend features. Then the probabilistic self-attention mechanism and self-attention distillation mechanism are used to effectively capture the global dependencies and local patterns of the input sequence. According to the output of the encoder, the decoder is used to generate the runoff prediction results, which specifically include the following parts: S5.1. Probabilistic sparse self-attention mechanism: The self-attention feature map of Transformer is a long-tail distribution. The prominent dot product will cause the corresponding attention distribution to be far away from the uniform distribution. According to the above principle, Informer uses KL divergence to measure the distribution distance. The larger the relative entropy, the greater the probability that it is an important part. The sparsity measure of the i-th query vector query is The formula is as follows: (3) In the formula, the first term is the Log-Sum-Exp(LSE) of the query vector, is the i-th query vector, is the key vector matrix, To select the length of the random key vector K, is the jth value in the transposed matrix of the key matrix K, The scaling factor added to prevent the gradient from vanishing, is an exponential function; the second term is the arithmetic mean. Based on the above method, the probability sparse self-attention calculation formula is as follows: (4) In the formula, are matrices consisting of query vector, key vector and value vector respectively, The scaling factor is increased to prevent the gradient from disappearing. Softmax is the activation function. is the query vector obtained by the above probability sparse measurement method, is the transpose of the key matrix K; S5.2, self-attention distillation mechanism: distillation operation is used at each layer to extract dominant features and generate compact self-attention feature maps at the next layer; the formula for the distillation process from layer j to layer j+1 is as follows: (5) In the formula, for The j+1 layer input after distillation, represents the attention block, represents the convolution operation, is the activation function, It is the maximum pooling operation; S5.3. Generative decoder structure: The standard decoding structure is adopted, which consists of two identical multi-head attention layers and generates prediction outputs through a feedforward network. The input vector of the decoder is as follows: (6) In the formula, is the t-th input sequence of the decoder, is the starting token of the t-th input sequence, It is a zero-valued placeholder for the target sequence. It is different from the self-attention mechanism in the encoder. Both of them process different types of input data. To set the data length range of the model.

[0042] In order to verify the effectiveness of STL decomposition of runoff signals for time series prediction models, the original data training data are: no runoff signal processing, runoff data processing using STL seasonal decomposition method, and runoff prediction by Informer model. The results are shown in Table 1.

[0043] Table 1. STL combination model parameter table

[0044] The results show that the use of STL has greatly improved the prediction accuracy of Informer, which proves that the introduction of STL into the Informer model can better capture and predict the changes in runoff and complete the prediction of runoff data.

[0045] In the preferred scheme, in Step 1, the selected monitoring power station range includes at least one hydropower station and its upstream rain gauge and hydrological station in the basin to improve the prediction accuracy; the above setting can comprehensively consider the changes in hydrological and meteorological conditions upstream of the hydropower station, and realize more accurate flood warning and power generation scheduling by real-time monitoring of the data of rain gauges and hydrological stations and combining the operation status of the hydropower station.

[0046] In the preferred scheme, in the Step 2, the specific implementation of the Pearson correlation analysis method includes calculating the Pearson correlation coefficient between each feature and the target prediction value, and selecting features whose absolute values ​​are greater than a preset threshold as strong correlation features, and the preset threshold is 0.4; the above settings can effectively screen out features with strong correlation with the target prediction value, reduce model complexity, and improve prediction accuracy. On the basis of the screened out strong correlation features, further feature selection and optimization are performed to lay a solid foundation for subsequent model training.

[0047] In the preferred solution, in Step 3, the STL seasonal decomposition method decomposes the original data into trend terms, seasonal terms and residual terms, where the decomposition period T is set to 365 days to ensure the accurate extraction of seasonal characteristics; the above settings enable the model to accurately capture the annual seasonal fluctuations of the data, thereby comprehensively improving the prediction accuracy of the overall time series data.

[0048] In the preferred solution, the specific steps of the Informer model modeling the input multi-source time series features through the encoder include: using the encoder in the model to model the input time series data and additional seasonal and trend features, and then using the probabilistic self-attention mechanism and the global-local information fusion mechanism to effectively capture the global dependencies and local patterns of the input sequence; according to the output of the encoder, using the decoder to generate the runoff prediction results; the above settings can significantly improve the model's ability to handle complex time series data and enhance the accuracy of the prediction. At the same time, by optimizing the loss function and training strategy, the prediction error is further reduced to ensure the stability and reliability of the model in practical applications.

[0049] The preferred solution also includes adjusting the water resource management and scheduling strategies of the hydropower station according to the prediction results to optimize the power generation plan and water storage management; the above settings are aimed at improving energy output efficiency, reducing water resource waste, and ensuring a stable power supply in drought or flood seasons, while protecting the downstream ecological and people's livelihood water needs.

[0050] In summary, the runoff prediction method based on multi-source information integration of the informer model provided by the present invention solves the problem of poor accuracy of medium- and long-term runoff prediction in the field of runoff prediction technology, especially for the prediction of runoff series with multi-source, nonlinear and non-stationary characteristics, and overcomes the limitation of poor prediction results of original long sequence data in the prior art; the multi-source information integration method based on the informer model proposed by the present invention is used for runoff prediction, which not only relies on the historical runoff data of the hydropower station, but also integrates the hydrological information such as rainfall and temperature in the surrounding area, thereby improving the accuracy and comprehensiveness of the prediction; it uses Pearson correlation analysis to select the data with strong correlation with the target prediction value The features of are used as the multi-source time series feature input of the Informer model, which enhances the feature capture ability of the model; the STL seasonal decomposition method is used to process the historical runoff information, rainfall and temperature information, and the periodic term and trend term features based on the historical series are extracted through feature expansion, which effectively identifies and analyzes the different components in the time series (such as trend, seasonality and residual), which helps to understand the basic structure and law of the data; through STL decomposition, the non-stationary characteristics of time series data can be reduced and the prediction accuracy of the model can be improved; the present invention innovatively applies the Informer model to runoff prediction, and uses the sparse attention mechanism and self-attention distillation mechanism of the model for efficient prediction; through the Informer The encoder of the model models the input time series data and the additional seasonal and trend features, and the decoder generates the prediction results of runoff, thereby realizing the accurate prediction of runoff at future moments. Through Pearson correlation analysis, the scheme can screen out the features with strong correlation with the target prediction value and use them as the multi-source time series feature input of the Informer model. This method not only improves the prediction accuracy of the model, but also reduces the unnecessary data processing burden. The present invention creatively combines the STL seasonal decomposition method with the Informer model, pre-processes the original data through STL decomposition, extracts the periodic term and trend term features, and uses them as the input of the Informer model. The method effectively utilizes the advantages of STL decomposition in time series analysis and the ability of the Informer model in processing long sequence data, thereby improving the accuracy and efficiency of runoff prediction; through the sparse attention mechanism and self-attention distillation mechanism of the Informer model, the scheme can efficiently capture the global dependencies and local patterns of the input sequence, and realize accurate prediction of runoff at future moments. This prediction method not only improves the prediction speed, but also enhances the model's adaptability to complex hydrological conditions; the present invention further improves the generalization ability and robustness of the model in runoff prediction by fusing feature information of different time scales, providing new technical ideas and practical paths for the field of hydrological forecasting.

Claims

1. A runoff prediction method based on multi-source information integration of the informer model, characterized in that: The method comprises the following steps: Step 1: Select the monitoring power station range and obtain the historical runoff data of the hydropower station and the hydrological information such as rainfall and temperature in the surrounding area; Step 2: Using the Pearson correlation analysis method, the features that are strongly correlated with the historical runoff data of the hydropower station are selected from the acquired hydrological information as the multi-source time series feature input of the informer model; Step 3: Use the STL seasonal decomposition method to preprocess the selected data set and extract the characteristics of periodic items and trend items; Step 4: Use the informer model to predict the runoff of the preprocessed data using the sparse attention mechanism and self-attention distillation mechanism to obtain the predicted runoff value at the future moment.

2. The runoff prediction method based on multi-source information integration of the informer model according to claim 1 is characterized by: In the Step 1, the selected monitoring power station range includes at least one hydropower station and a rainfall station and a hydrological station in its upstream basin to improve the prediction accuracy.

3. The runoff prediction method based on multi-source information integration of the informer model according to claim 1 is characterized by: In the Step 2, the specific implementation of the Pearson correlation analysis method includes calculating the Pearson correlation coefficient between each feature and the target prediction value, and selecting features whose absolute values ​​are greater than a preset threshold as strong correlation features, and the preset threshold is 0.

4.

4. The runoff prediction method based on multi-source information integration of the informer model according to claim 3 is characterized by: The calculation formula of the Pearson correlation coefficient is as follows: （1）； In the formula, is the i-th data of the j-th data to be predicted, is the jth need and historical data The i-th data of the hydrological information to be analyzed, and They are and The mean of yes and The correlation coefficient of .

5. The runoff prediction method based on multi-source information integration of the informer model according to claim 1 is characterized by: In Step 3, the STL seasonal decomposition method decomposes the original data into trend terms, seasonal terms and residual terms, wherein the decomposition period T is set to 365 days to ensure accurate extraction of seasonal features.

6. The runoff prediction method based on multi-source information integration of the informer model according to claim 5 is characterized by: The detailed steps of the STL seasonal decomposition method include: Step 3.1: Divide the original data into multiple subsequences according to the size of the decomposition period, and then perform local polynomial smoothing on the subsequences to remove noise and rapidly changing parts to obtain a preliminary trend estimate; Step 3.2: Calculate the seasonal average at each time to estimate the seasonal component, remove the trend and seasonal data by subtraction, and perform smoothing again to obtain the residual component; Step 3.3: Repeat the decomposition several times until the estimated values ​​of the trend term and seasonal term converge.

7. The runoff prediction method based on multi-source information integration of the informer model according to claim 1 is characterized by: In the Step 4, the informer model models the input multi-source time series features through the encoder, generates runoff prediction results using the decoder, and improves the prediction speed and accuracy through the self-attention distillation mechanism.

8. The runoff prediction method based on multi-source information integration of the informer model according to claim 7 is characterized by: The specific steps of the informer model for modeling the input multi-source time series features through the encoder include: using the encoder in the model to model the input time series data and additional seasonal and trend features, and then using the probabilistic self-attention mechanism and the global-local information fusion mechanism to effectively capture the global dependencies and local patterns of the input sequence; according to the output of the encoder, using the decoder to generate the runoff prediction result.

9. The runoff prediction method based on multi-source information integration of the informer model according to claim 1 is characterized by: It also includes adjusting the water resource management and scheduling strategies of hydropower stations based on the forecast results to optimize power generation plans and water storage management.

Images