Method and system for medium and long term runoff prediction based on multi-source climate factors and deep learning, electronic device and storage medium

Abstract

The application discloses a method, system, device and medium for medium and long term runoff prediction based on multi-source climate factors and deep learning. The method comprises: obtaining historical runoff time series data and multi-source climate factor data, wherein the multi-source climate factors are divided into three-dimensional space climate field factors and two-dimensional climate index factors; based on an initial space window scale, the three-dimensional space climate field factors are cropped and traversed at different space positions, a runoff prediction model is trained in a first stage, and candidate space driving region centers are determined by comparing prediction performance; taking the candidate space driving region centers as space window centers, a plurality of different space window scales are traversed, a runoff prediction model is trained in a second stage, the optimal space window scale corresponding to each candidate space driving region center and the final parameters of the model are determined by prediction performance, a plurality of candidate runoff prediction models are constructed, and the optimal model for final runoff prediction is determined by comprehensive comparison. The application can improve the prediction accuracy, stability and reliability of runoff prediction.

Description

Technical Field

[0001] This invention relates to the field of hydrological forecasting technology, specifically to a medium- and long-term runoff prediction method, system, electronic device, and storage medium based on multi-source climate factors and deep learning. Background Technology

[0002] Medium- and long-term runoff forecasting is a crucial technical support for watershed water resource allocation, flood control and disaster reduction, and hydrological planning. However, runoff variation is influenced by a variety of climate factors, including ocean-atmosphere coupling processes, large-scale circulation anomalies, and polar environmental changes. In particular, on medium- and long-term timescales, runoff variation is not only affected by local precipitation and temperature conditions, but also closely related to long-range climate factors such as sea surface temperature anomalies, atmospheric circulation patterns, and polar sea ice changes.

[0003] Existing medium- and long-term runoff forecasting methods mainly include statistical modeling methods and forecasting methods based on machine learning or deep learning. Traditional statistical methods mostly construct forecasting relationships based on correlation analysis, regression models, or empirical formulas, which are difficult to characterize the complex nonlinear coupling relationships between multiple climate factors, and lack forecast stability and adaptability when the climate background changes.

[0004] In recent years, deep learning-based runoff prediction methods have made some progress in simulation accuracy, but they still generally suffer from insufficient identification of spatial climate-driven regions. Existing methods often directly use fixed regional or global climate fields as model inputs, failing to systematically identify key spatial climate-driven regions that play a dominant role in runoff changes in the target watershed. This results in a large amount of spatially redundant information in the model input, which not only increases the complexity of model training but may also weaken the ability to characterize key signals, thereby affecting prediction accuracy and model generalization performance.

[0005] Furthermore, existing methods typically lack an integrated process for identifying spatial driving zones and optimizing spatial window scales based on prediction performance feedback, making it difficult to automatically identify key spatial climate driving zones and optimize their spatial extent. For different watersheds, the spatial range of climate driving signals usually varies. Directly using a fixed spatial range as model input can easily reduce the matching degree between the model input and the watershed runoff response characteristics, thereby further affecting the stability and reliability of medium- and long-term runoff prediction results.

[0006] Therefore, there is an urgent need for a method that can systematically identify key spatial climate-driven regions and optimize the configuration of spatial window scales in medium- and long-term runoff forecasting, so as to improve the effectiveness of model inputs and the stability and reliability of forecast results while ensuring forecast accuracy. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a medium- to long-term runoff prediction method, system, electronic device, and storage medium based on multi-source climate factors and deep learning, applicable to runoff prediction at watershed, monthly, and longer time scales.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A medium- to long-term runoff prediction method based on multi-source climate factors and deep learning includes the following steps: S1: Obtain historical runoff time series data of the target watershed and corresponding multi-source climate factor data to construct an initial dataset for runoff prediction; the multi-source climate factor data is divided into three-dimensional spatial climate field factors and two-dimensional climate index factors. S2: Based on the initial dataset, determine the initial spatial window scale for locating the center of the candidate spatial driving zone. Under this initial spatial window scale, perform full spatial location traversal and pruning of the three-dimensional spatial climate field factors, construct the first round of runoff prediction models, and train and predict runoff. Determine the center of the candidate spatial driving zone based on the comparison results of prediction performance. S3: Using the location of each candidate spatial driving center as the center of the spatial window, traverse multiple different spatial window scales, construct the second round of runoff prediction model and predict runoff, compare the performance of the prediction results, determine the optimal spatial window scale corresponding to each candidate spatial driving center, and obtain several candidate runoff prediction models that correspond one-to-one with the spatial driving center and the spatial window scale. S4: Conduct a comprehensive comparison of the prediction performance of each candidate runoff prediction model to determine the optimal model for the final runoff prediction; S5: Input the real-time three-dimensional spatial climate field factors and real-time two-dimensional climate index factors within the optimal spatial window determined by the optimal spatial driving center and the optimal spatial window scale into the optimal model to predict medium- and long-term runoff.

[0010] To optimize the above technical solution, the specific measures also include: Furthermore, in S1, the three-dimensional spatial climate field factors include the sea surface temperature field, the atmospheric geopotential height field, and the westerly-related spatial field; the two-dimensional climate index factors include the polar sea ice area index, the El Niño-Southern Oscillation index, and the North Atlantic Oscillation index. The construction of the initial dataset for runoff prediction specifically involves preprocessing historical runoff time series data and the multi-source climate factor data, including time alignment, missing data processing, and standardization, to form the initial dataset.

[0011] Furthermore, S2 specifically includes: S2.1 Construct the model input for the first round of runoff prediction simulation, wherein the model input includes a spatial climate field input obtained by spatially cropping the three-dimensional spatial climate field factors at the initial spatial window scale, and a two-dimensional climate index factor corresponding to the time of the spatial climate field input. S2.2 Based on the initial data set constructed in step S1, and combined with the model input for the first round of runoff prediction simulation and the corresponding historical runoff data, construct a sample dataset for the first round of runoff prediction, and divide the sample dataset for the first round of runoff prediction into a first training dataset and a first validation dataset. S2.3 Construct the first round of runoff prediction model, input the first training dataset into the first round of runoff prediction model to perform the first round of runoff prediction simulation, and output the corresponding runoff prediction results; the first round of runoff prediction model includes a feature extraction module and a time series modeling module; the feature extraction module is used to extract the spatial features of three-dimensional spatial climate field factors, and the time series modeling module is used to model the time dependence between multi-source climate factors and runoff; S2.4. Based on the first validation dataset, compare the prediction performance of the first round of runoff prediction results corresponding to different spatial locations. The prediction performance is measured by at least one of the following evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Nash efficiency coefficient. Based on the prediction performance ranking results, select spatial locations with prediction performance better than a preset threshold or ranked in the top N as candidate spatial driving zone centers.

[0012] Furthermore, S3 specifically includes: S3.1 For each candidate spatial driving zone center, take the location of the candidate spatial driving zone center as the spatial window center, construct multiple spatial windows of different scales respectively, and use the multiple spatial windows of different scales to spatially clip the three-dimensional spatial climate field factors to obtain multiple spatial climate field inputs, and combine the spatial climate field inputs with the two-dimensional climate index factors corresponding to their time to form the second round of runoff prediction model inputs. S3.2 Based on the initial data set constructed in step S1, and combined with the input of the second round of runoff prediction model and the corresponding historical runoff data, construct a sample dataset for the second round of runoff prediction, and divide the sample dataset for the second round of runoff prediction into a second training dataset, a second validation dataset and a second test dataset. S3.3 Construct the second runoff prediction model. The structure of the second runoff prediction model is the same as that of the first runoff prediction model. Input the second training dataset into the second runoff prediction model to simulate the second runoff prediction and obtain the second runoff prediction results under different spatial window scales. S3.4 Based on the second validation dataset, the performance of the second-round runoff prediction results corresponding to different spatial window scales under the same candidate spatial driving zone center is compared. The prediction performance is measured by at least one of the following evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Nash efficiency coefficient. The different spatial window scales are sorted according to the prediction performance evaluation results to determine the optimal spatial window scale corresponding to the candidate spatial driving zone center. The final parameters of the second-round runoff prediction model under the optimal spatial window scale and its corresponding model input conditions are obtained.

[0013] Furthermore, S4 specifically refers to: Based on the centers of different candidate spatial driving zones, the optimal spatial window scale corresponding to each candidate spatial driving zone center, and the final parameters of the second-round runoff prediction model, multiple candidate runoff prediction models are constructed. The prediction performance of each candidate runoff prediction model is evaluated using the second test dataset obtained in step S3.2. The prediction performance is measured by at least one of the following evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Nash efficiency coefficient. The model with the best prediction performance is determined from the multiple candidate runoff prediction models, and the final spatial driving zone center location and spatial window scale are uniquely determined by the optimal model.

[0014] Furthermore, the feature extraction module employs a convolutional neural network or an encoder structure; the time series modeling module employs a long short-term memory network or a recurrent neural network.

[0015] This invention also proposes a medium- to long-term runoff prediction system based on multi-source climate factors and deep learning, comprising: The initial data acquisition module is used to acquire historical runoff time series data of the target watershed and corresponding multi-source climate factor data to construct an initial dataset for runoff prediction; the multi-source climate factor data is divided into three-dimensional spatial climate field factors and two-dimensional climate index factors. The driving zone center localization module is used to determine the initial spatial window scale for candidate spatial driving zone center localization based on the initial dataset. Under the initial spatial window scale, the three-dimensional spatial climate field factors are traversed and clipped in full space. The first round of runoff prediction models are constructed and trained and predicted for runoff. The candidate spatial driving zone centers are determined by comparing the prediction performance. The window scale optimization module is used to traverse multiple different spatial window scales, using the location of each candidate spatial driving center as the spatial window center, to construct a second-round runoff prediction model and predict runoff. The performance of the prediction results is compared to determine the optimal spatial window scale corresponding to each candidate spatial driving center, and several candidate runoff prediction models that correspond one-to-one with the spatial driving center and the spatial window scale are obtained. The model selection module is used to comprehensively compare the prediction performance of each candidate runoff prediction model and determine the optimal model for the final runoff prediction. The prediction module is used to input real-time three-dimensional spatial climate field factors and real-time two-dimensional climate index factors within the optimal spatial window determined by the optimal spatial driving center and the optimal spatial window scale into the optimal model to predict medium- and long-term runoff.

[0016] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described above.

[0017] The present invention also proposes a computer-readable storage medium storing a computer program that enables a computer to execute the medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described above.

[0018] The beneficial effects of this invention are: (1) The spatial driving area has strong objectivity in identification. This invention objectively determines the key spatial regions that play a dominant role in the runoff change of the target watershed by performing full-space traversal and clipping of spatial climate field factors under a preset initial spatial window scale and identifying candidate spatial driving zone centers based on prediction performance feedback, thus avoiding the uncertainty brought about by empirical selection.

[0019] (2) The spatial scale and watershed response characteristics are reasonably matched. This invention uses the center of the candidate spatial driving region as the window center, traverses multiple different spatial window scales, and compares and ranks the prediction performance of models corresponding to different spatial window scales based on the validation dataset, thereby achieving adaptive matching between spatial window scale and watershed runoff response characteristics.

[0020] (3) The process of determining the optimal model is clear and the results are stable. This invention constructs multiple candidate runoff prediction models by combining the center of the candidate spatial driving region, its corresponding optimal spatial window scale, and the final parameters of the model. It then uses a test dataset to conduct a unified evaluation and comprehensive comparison of each candidate model, thereby determining the optimal model for the final runoff prediction and improving the stability and reliability of medium- and long-term runoff prediction results. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the medium- and long-term runoff prediction method based on multi-source climate factors and deep learning proposed in this invention.

[0022] Figure 2This is a schematic diagram showing the location of the Ertis River basin and the distribution of hydrological stations selected in this invention.

[0023] Figure 3 This is a schematic diagram showing the distribution of predictive performance at different spatial locations of each hydrological station in the 500 hPa height field (HGT).

[0024] Figure 4 This is a schematic diagram showing the distribution of predictive performance of sea surface temperature (SST) at different spatial locations for each hydrological station.

[0025] Figure 5 This is a schematic diagram showing the distribution of predictive performance of each hydrological station at different spatial locations corresponding to the Westerly Wind Index (WI).

[0026] Figure 6 This diagram illustrates the comparison of model prediction performance at different spatial window scales.

[0027] Figure 7 This diagram illustrates the comparison between the predicted results of the final runoff prediction model and the measured runoff.

[0028] Figure 8 A schematic diagram showing the contribution ratio of different climate factors to runoff prediction.

[0029] Figure 9 This is a schematic diagram illustrating the nonlinear response characteristics of climate factors to runoff prediction results within different value ranges. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0031] Example 1 This embodiment takes the Ertis River Basin as the research object and selects historical monthly runoff time series data of typical hydrological stations in the basin as the prediction target variable.

[0032] This invention proposes a medium- to long-term runoff prediction method based on multi-source climate factors and deep learning. The process of this method is as follows: Figure 1 As shown, it includes the following steps: S1: Obtain historical runoff time series data and corresponding multi-source climate factor data for the target watershed to construct an initial dataset for runoff prediction; the multi-source climate factor data is divided into three-dimensional spatial climate field factors and two-dimensional climate index factors; the three-dimensional spatial climate field factors include sea surface temperature field, atmospheric geopotential height field and westerly-related spatial field; the two-dimensional climate index factors include polar sea ice area index, El Niño-Southern Oscillation index and North Atlantic Oscillation index; In this embodiment, the multi-source climate factors include: three-dimensional spatial climate field factors with spatial distribution characteristics, namely Atlantic sea surface temperature (SST), 500 hPa geopotential height (HGT), and westerly index (WI); and two-dimensional climate index factors without spatial distribution but changing only over time, namely Arctic sea ice area (NSI), El Niño-Southern Oscillation Index (ENSO), and North Atlantic Oscillation Index (NAO). The above runoff data and multi-source climate factor data are time-aligned, have missing data processed, and standardized to construct an initial dataset for runoff prediction modeling. Figure 2 This is a schematic diagram showing the location of the Ertis River basin and the distribution of hydrological stations selected in this invention.

[0033] S2: Based on the initial dataset, determine the initial spatial window scale for locating the center of the candidate spatial driving region. The initial spatial window scale is a preset spatial window scale, and its corresponding spatial range is smaller than the minimum scale range in the subsequent spatial window scale optimization stage, so as to improve the spatial resolution of the candidate spatial driving region center identification. Under this initial spatial window scale, the three-dimensional spatial climate field factors are traversed and pruned in all spatial locations to construct the first round of runoff prediction models and conduct training and runoff prediction; the candidate spatial driving region centers are determined by comparing the prediction performance results; S2 specifically includes: S2.1 Construct the model input for the first round of runoff prediction simulation, wherein the model input includes a spatial climate field input obtained by spatially cropping the three-dimensional spatial climate field factors at the initial spatial window scale, and a two-dimensional climate index factor corresponding to the time of the spatial climate field input. S2.2 Based on the initial data set constructed in step S1, and combined with the model input for the first round of runoff prediction simulation and the corresponding historical runoff data, construct a sample dataset for the first round of runoff prediction, and divide the sample dataset for the first round of runoff prediction into a first training dataset and a first validation dataset. S2.3. Construct the first-round runoff prediction model, input the first training dataset into the first-round runoff prediction model to perform the first-round runoff prediction simulation, and output the corresponding runoff prediction results; the first-round runoff prediction model includes a feature extraction module and a time series modeling module; the feature extraction module is used to extract the spatial features of three-dimensional spatial climate field factors, and the time series modeling module is used to model the time dependence between multi-source climate factors and runoff; the feature extraction module can use convolutional neural networks, encoder structures or other spatial feature extraction models; the time series modeling module can use long short-term memory networks, recurrent neural networks or other time series modeling models.

[0034] S2.4. Based on the first validation dataset, compare the prediction performance of the first round of runoff prediction results corresponding to different spatial locations. The prediction performance is measured by at least one of the following evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Nash efficiency coefficient. Based on the prediction performance ranking results, select spatial locations with prediction performance better than a preset threshold or ranked in the top N as candidate spatial driving zone centers.

[0035] See the schematic diagram of the predicted performance distribution of HGT at different spatial locations of each hydrological station. Figure 3 See the schematic diagram of the predictive performance distribution of SST at different spatial locations of various hydrological stations. Figure 4 See the schematic diagram of the prediction performance distribution of each hydrological station at different spatial locations. Figure 5 .

[0036] S3: Using the location of each candidate spatial driving center as the center of the spatial window, traverse multiple different spatial window scales to construct a second-round runoff prediction model and predict runoff. Compare the performance of the prediction results to determine the optimal spatial window scale corresponding to each candidate spatial driving center, resulting in several candidate runoff prediction models that correspond one-to-one with the spatial driving center and the spatial window scale. S3 specifically includes: S3.1 For each candidate spatial driving zone center, take the location of the candidate spatial driving zone center as the spatial window center, construct multiple spatial windows of different scales respectively, and use the multiple spatial windows of different scales to spatially clip the three-dimensional spatial climate field factors to obtain multiple spatial climate field inputs, and combine the spatial climate field inputs with the two-dimensional climate index factors corresponding to their time to form the second round of runoff prediction model inputs. S3.2 Based on the initial data set constructed in step S1, and combined with the input of the second round of runoff prediction model and the corresponding historical runoff data, construct a sample dataset for the second round of runoff prediction, and divide the sample dataset for the second round of runoff prediction into a second training dataset, a second validation dataset and a second test dataset. S3.3 Construct the second runoff prediction model. The structure of the second runoff prediction model is the same as that of the first runoff prediction model. Input the second training dataset into the second runoff prediction model to simulate the second runoff prediction and obtain the second runoff prediction results under different spatial window scales. S3.4 Based on the second validation dataset, the performance of the second-round runoff prediction results corresponding to different spatial window scales under the same candidate spatial driving zone center is compared. The prediction performance is measured by at least one of the following evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Nash efficiency coefficient. The different spatial window scales are sorted according to the prediction performance evaluation results to determine the optimal spatial window scale corresponding to the candidate spatial driving zone center. The final parameters of the second-round runoff prediction model under the optimal spatial window scale and its corresponding model input conditions are obtained.

[0037] See the diagram comparing the model prediction performance at different spatial window scales. Figure 6 .

[0038] S4: A comprehensive comparison of the prediction performance of each candidate runoff prediction model is performed to determine the optimal model for the final runoff prediction; S4 specifically involves: Based on the centers of different candidate spatial driving zones, the optimal spatial window scale corresponding to each candidate spatial driving zone center, and the final parameters of the second-round runoff prediction model, multiple candidate runoff prediction models are constructed. The prediction performance of each candidate runoff prediction model is evaluated using the second test dataset obtained in step S3.2. The prediction performance is measured by at least one of the following evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Nash efficiency coefficient. The model with the best prediction performance is determined from the multiple candidate runoff prediction models, and the final spatial driving zone center location and spatial window scale are uniquely determined by the optimal model.

[0039] S5: Input the real-time three-dimensional spatial climate field factors and real-time two-dimensional climate index factors within the optimal spatial window determined by the optimal spatial driving center and the optimal spatial window scale into the optimal model to predict medium- and long-term runoff.

[0040] Compared to methods that directly use the global climate field as model input, this invention, through a two-stage spatial driving zone localization and spatial window scale optimization mechanism, can reduce the interference of spatial redundancy information on model training and improve the matching degree between model input and watershed runoff response characteristics, under the same model structure conditions. This is beneficial to improving runoff prediction accuracy and model stability. See the schematic diagram comparing the final model prediction results with measured runoff. Figure 7 .

[0041] Analysis of the driving mechanism of this invention: After determining the optimal model for final runoff prediction, a surrogate model is constructed based on the model input data and corresponding runoff prediction output data of the optimal model to approximate the input-output mapping relationship of the final model. The surrogate model learns the prediction response characteristics of the optimal model under multi-dimensional input conditions, thereby achieving a computable expression of the prediction behavior of the final model.

[0042] The model input of the optimal model is mapped to an interpretable feature space, which is obtained by statistically summarizing and structurally representing multi-source climate factors. Based on the surrogate model, the marginal contribution of each interpretable feature (multi-source climate factor) to the runoff prediction result under given input sample conditions is calculated in the interpretable feature space to quantify the degree of influence of different interpretable features on the prediction result.

[0043] The contribution values ​​are aggregated according to the climatic factor categories to obtain the overall contribution value of each climatic factor in the runoff prediction process. By determining the positive or negative sign and magnitude of the contribution values, it is possible to distinguish whether each climatic factor has a promoting or inhibiting effect on runoff changes.

[0044] Meanwhile, based on the correspondence between interpretable feature contribution values ​​and the original values ​​of corresponding climate factors, by analyzing the trend of contribution value changes corresponding to different climate factor value ranges, the ranges of promoting or inhibiting effects of key climate factors on runoff prediction results are identified. This reveals the model's response characteristics under different climate background conditions, clarifies the model's decision-making basis under different runoff scenarios, and enhances the interpretability and credibility of medium- and long-term runoff prediction results. When climate factors are in the low-intensity range, their contribution values ​​are close to zero; when climate factors exceed a certain threshold range, their contribution values ​​significantly increase or undergo a sign change, thus exhibiting obvious threshold-type nonlinear response characteristics. As a preferred approach, the weights of each climate factor in the dataset used to train the runoff prediction model can be adjusted based on the contribution values ​​of the climate factors, increasing the weight of climate factors with high contribution values ​​and decreasing the weight of climate factors with low contribution values.

[0045] By interpreting and analyzing the driving mechanisms, this invention can identify the dominant role of different climate factors in runoff changes at different time stages, reveal the intrinsic basis of model prediction decisions, and thus enhance the interpretability and decision-making reference value of medium- and long-term runoff prediction results.

[0046] See the diagram illustrating the contribution ratio of different climate factors to runoff prediction. Figure 8 For the nonlinear response characteristics of climate factors to runoff prediction in different value ranges, please refer to [link / reference]. Figure 9 .

[0047] Example 2 This invention proposes a medium- to long-term runoff prediction system based on multi-source climate factors and deep learning, corresponding to the method in Example 1, comprising: The initial data acquisition module is used to acquire historical runoff time series data of the target watershed and corresponding multi-source climate factor data to construct an initial dataset for runoff prediction; the multi-source climate factor data is divided into three-dimensional spatial climate field factors and two-dimensional climate index factors. The driving zone center localization module is used to determine the initial spatial window scale for candidate spatial driving zone center localization based on the initial dataset. Under the initial spatial window scale, the three-dimensional spatial climate field factors are traversed and clipped in full space. The first round of runoff prediction models are constructed and trained and predicted for runoff. The candidate spatial driving zone centers are determined by comparing the prediction performance. The window scale optimization module is used to traverse multiple different spatial window scales, using the location of each candidate spatial driving center as the spatial window center, to construct a second-round runoff prediction model and predict runoff. The performance of the prediction results is compared to determine the optimal spatial window scale corresponding to each candidate spatial driving center, and several candidate runoff prediction models that correspond one-to-one with the spatial driving center and the spatial window scale are obtained. The model selection module is used to comprehensively compare the prediction performance of each candidate runoff prediction model and determine the optimal model for the final runoff prediction. The prediction module is used to input real-time three-dimensional spatial climate field factors and real-time two-dimensional climate index factors within the optimal spatial window determined by the optimal spatial driving center and the optimal spatial window scale into the optimal model to predict medium- and long-term runoff.

[0048] The implementation methods of each module and its function in the system are completely consistent with the steps of the method in Implementation Example 1, so they will not be repeated here.

[0049] Example 3 This invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described in Embodiment 1.

[0050] Example 4 This invention proposes a computer-readable storage medium storing a computer program that enables a computer to execute a medium- to long-term runoff prediction method based on multi-source climate factors and deep learning as described in Example 1.

[0051] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, and portable compact disc read-only memory (CD). ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0052] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0053] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A medium- to long-term runoff prediction method based on multi-source climate factors and deep learning, characterized in that, Includes the following steps: S1: Obtain historical runoff time series data of the target watershed and corresponding multi-source climate factor data to construct an initial dataset for runoff prediction; the multi-source climate factor data is divided into three-dimensional spatial climate field factors and two-dimensional climate index factors. S2: Based on the initial dataset, determine the initial spatial window scale for locating the center of the candidate spatial driving zone. Under this initial spatial window scale, perform full spatial location traversal and pruning of the three-dimensional spatial climate field factors, construct the first round of runoff prediction models, and train and predict runoff. Determine the center of the candidate spatial driving zone based on the comparison results of prediction performance. S3: Using the location of each candidate spatial driving center as the center of the spatial window, traverse multiple different spatial window scales, construct the second round of runoff prediction model and predict runoff, compare the performance of the prediction results, determine the optimal spatial window scale corresponding to each candidate spatial driving center, and obtain several candidate runoff prediction models that correspond one-to-one with the spatial driving center and the spatial window scale. S4: Conduct a comprehensive comparison of the prediction performance of each candidate runoff prediction model to determine the optimal model for the final runoff prediction; S5: Input the real-time three-dimensional spatial climate field factors and real-time two-dimensional climate index factors within the optimal spatial window determined by the optimal spatial driving center and the optimal spatial window scale into the optimal model to predict medium- and long-term runoff.

2. The medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described in claim 1, characterized in that, In S1, the three-dimensional spatial climate field factors include the sea surface temperature field, the atmospheric geopotential height field, and the westerly-related spatial field; the two-dimensional climate index factors include the polar sea ice area index, the El Niño-Southern Oscillation index, and the North Atlantic Oscillation index. The construction of the initial dataset for runoff prediction specifically involves preprocessing historical runoff time series data and the multi-source climate factor data, including time alignment, missing data processing, and standardization, to form the initial dataset.

3. The medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described in claim 1, characterized in that, S2 specifically includes: S2.1 Construct the model input for the first round of runoff prediction simulation, wherein the model input includes a spatial climate field input obtained by spatially cropping the three-dimensional spatial climate field factors at the initial spatial window scale, and a two-dimensional climate index factor corresponding to the time of the spatial climate field input. S2.2 Based on the initial data set constructed in step S1, and combined with the model input for the first round of runoff prediction simulation and the corresponding historical runoff data, construct a sample dataset for the first round of runoff prediction, and divide the sample dataset for the first round of runoff prediction into a first training dataset and a first validation dataset. S2.3 Construct the first round of runoff prediction model, input the first training dataset into the first round of runoff prediction model to perform the first round of runoff prediction simulation, and output the corresponding runoff prediction results; the first round of runoff prediction model includes a feature extraction module and a time series modeling module; the feature extraction module is used to extract the spatial features of three-dimensional spatial climate field factors, and the time series modeling module is used to model the time dependence between multi-source climate factors and runoff; S2.

4. Based on the first validation dataset, compare the prediction performance of the first round of runoff prediction results corresponding to different spatial locations. The prediction performance is measured by at least one of the following evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Nash efficiency coefficient. Based on the prediction performance ranking results, select spatial locations with prediction performance better than a preset threshold or ranked in the top N as candidate spatial driving zone centers.

4. The medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described in claim 1, characterized in that, S3 specifically includes: S3.1 For each candidate spatial driving zone center, take the location of the candidate spatial driving zone center as the spatial window center, construct multiple spatial windows of different scales respectively, and use the multiple spatial windows of different scales to spatially clip the three-dimensional spatial climate field factors to obtain multiple spatial climate field inputs, and combine the spatial climate field inputs with the two-dimensional climate index factors corresponding to their time to form the second round of runoff prediction model inputs. S3.2 Based on the initial data set constructed in step S1, and combined with the input of the second round of runoff prediction model and the corresponding historical runoff data, construct a sample dataset for the second round of runoff prediction, and divide the sample dataset for the second round of runoff prediction into a second training dataset, a second validation dataset and a second test dataset. S3.3 Construct the second runoff prediction model. The structure of the second runoff prediction model is the same as that of the first runoff prediction model. Input the second training dataset into the second runoff prediction model to simulate the second runoff prediction and obtain the second runoff prediction results under different spatial window scales. S3.4 Based on the second validation dataset, the performance of the second-round runoff prediction results corresponding to different spatial window scales under the same candidate spatial driving zone center is compared. The prediction performance is measured by at least one of the following evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Nash efficiency coefficient. The different spatial window scales are sorted according to the prediction performance evaluation results to determine the optimal spatial window scale corresponding to the candidate spatial driving zone center. The final parameters of the second-round runoff prediction model under the optimal spatial window scale and its corresponding model input conditions are obtained.

5. The medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described in claim 4, characterized in that, S4 specifically refers to: Based on the centers of different candidate spatial driving zones, the optimal spatial window scale corresponding to each candidate spatial driving zone center, and the final parameters of the second-round runoff prediction model, multiple candidate runoff prediction models are constructed. The prediction performance of each candidate runoff prediction model is evaluated using the second test dataset obtained in step S3.

2. The prediction performance is measured by at least one of the following evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Nash efficiency coefficient. The model with the best prediction performance is determined from the multiple candidate runoff prediction models, and the final spatial driving zone center location and spatial window scale are uniquely determined by the optimal model.

6. The medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described in claim 3, characterized in that, The feature extraction module employs a convolutional neural network or encoder structure; the time series modeling module employs a long short-term memory network or a recurrent neural network.

7. A medium- to long-term runoff prediction system based on multi-source climate factors and deep learning, characterized in that, include: The initial data acquisition module is used to acquire historical runoff time series data of the target watershed and corresponding multi-source climate factor data to construct an initial dataset for runoff prediction. Multi-source climate factor data are divided into three-dimensional spatial climate field factors and two-dimensional climate index factors; The driving zone center localization module is used to determine the initial spatial window scale for candidate spatial driving zone center localization based on the initial dataset. Under the initial spatial window scale, the three-dimensional spatial climate field factors are traversed and clipped in full space. The first round of runoff prediction models are constructed and trained and predicted for runoff. The candidate spatial driving zone centers are determined by comparing the prediction performance. The window scale optimization module is used to traverse multiple different spatial window scales, using the location of each candidate spatial driving center as the spatial window center, to construct a second-round runoff prediction model and predict runoff. The performance of the prediction results is compared to determine the optimal spatial window scale corresponding to each candidate spatial driving center, and several candidate runoff prediction models that correspond one-to-one with the spatial driving center and the spatial window scale are obtained. The model selection module is used to comprehensively compare the prediction performance of each candidate runoff prediction model and determine the optimal model for the final runoff prediction. The prediction module is used to input real-time three-dimensional spatial climate field factors and real-time two-dimensional climate index factors within the optimal spatial window determined by the optimal spatial driving center and the optimal spatial window scale into the optimal model to predict medium- and long-term runoff.

8. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program enables the computer to execute the medium- and long-term runoff prediction method based on multi-source climate factors and deep learning as described in any one of claims 1-6.

Images