Multi-physical factor coupled watershed cascade hydropower station power generation prediction method and system

Abstract

This invention discloses a method and system for predicting the power generation of a cascade hydropower station in a river basin using multi-physical factor coupling, belonging to the field of hydropower technology. The prediction method includes acquiring and preprocessing multi-source data of the target area to construct a high-quality dataset; classifying the river basins within the target area into different types based on hydropower distribution characteristics; constructing multi-physical factor feature vectors for different types of river basins based on the high-quality dataset, and performing feature filtering to obtain the optimal input feature set; constructing and training suitable prediction models for different types of river basins based on the optimal input feature set; inputting the multi-physical factor feature vectors for the target time period into the trained suitable prediction model, outputting and accumulating the predicted power generation values ​​for each river basin to obtain the total hydropower generation of the target area. This invention solves the problems of missing physical mechanisms, poor river basin adaptability, and difficulty in achieving both prediction accuracy and engineering application in existing long-term hydropower generation prediction technologies.

Description

Technical Field

[0001] This invention belongs to the field of hydropower technology, specifically relating to a method and system for predicting the power generation of a cascade hydropower station in a river basin, involving multiple physical factors coupled together. Background Technology

[0002] Hydropower, as a core pillar of clean and renewable energy, exhibits significant seasonality and randomness in its power generation, influenced by multiple factors including meteorological conditions, hydrological situations, and reservoir scheduling strategies. Medium- to long-term power generation forecasts (1 to 6 months) are crucial for energy authorities to formulate power production plans, coordinate multi-energy complementarity (hydropower, thermal power, wind power, and solar power), and resolve the contradictions of wasted hydropower and idled electricity. They are of great significance for ensuring the stable operation of the power system and achieving the "dual carbon" target.

[0003] Current medium- and long-term forecasting technologies for the power generation of cascade hydropower stations in river basins still have three major shortcomings: First, there is a disconnect between physical mechanisms and data-driven approaches. Traditional hydrological models rely on basin-specific parameter calibration, making it difficult to achieve nationwide deployment; while purely data-driven machine learning models lack the physical constraints of hydrological cycles and reservoir scheduling, making them prone to predictive instability under extreme weather conditions.

[0004] Secondly, there is homogenization in watershed modeling. Existing technologies fail to distinguish between major watersheds with controlling hydropower stations and other watersheds dominated by small and medium-sized hydropower stations, ignoring the fundamental differences between the two in terms of hydrological response characteristics and dispatching and control capabilities, resulting in insufficient overall prediction accuracy.

[0005] Third, the project lacks applicability. There is a lack of technical solutions that balance forecast accuracy and operational efficiency, making it difficult to meet the operational needs of energy management departments to release monthly rolling forecasts of total hydropower generation.

[0006] To address the aforementioned issues, there is an urgent need for a predictive method that can deeply integrate physical mechanisms and data-driven approaches, and perform differentiated modeling for different types of watersheds, in order to fill the gaps in existing technologies. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for predicting the power generation of a cascade hydropower station in a river basin by coupling multiple physical factors. This method can solve the problems of lack of physical mechanisms, poor river basin adaptability, and difficulty in achieving both prediction accuracy and engineering application accuracy in existing long-term hydropower generation prediction technologies.

[0008] In a first aspect, the present invention provides a method for predicting the power generation of a cascade hydropower station in a river basin by coupling multiple physical factors, comprising the following steps: Acquire multi-source data from the target region and preprocess it to construct a high-quality dataset; Based on the characteristics of hydropower distribution, the watersheds within the target area are divided into different types; Based on high-quality datasets, multi-physical factor feature vectors are constructed for different types of watersheds, and feature selection is performed to obtain the optimal input feature set; Based on the optimal input feature set, we construct and train adaptive prediction models for different types of watersheds. The multi-physical factor feature vectors of the target time period are input into the trained and adapted prediction model, and the predicted values ​​of power generation in each watershed are output and accumulated to obtain the total hydropower generation of the target area.

[0009] As an alternative implementation, constructing a high-quality dataset further includes calculating the total effective rainfall of the watershed using a grid-based installed capacity weighting method; The grid-based installed capacity weighted method includes: calculating the grid installed capacity weight coefficient based on the installed capacity of hydropower stations within the grid; calculating the grid terrain correction coefficient based on the grid terrain features; calculating the grid effective rainfall by combining the grid measured rainfall, the grid installed capacity weight coefficient, and the grid terrain correction coefficient; and finally, obtaining the total effective rainfall of the watershed by weighting and accumulating the grid areas.

[0010] As an alternative implementation method, the watersheds within the target area are divided into different types: sub-watersheds that include control cascade hydropower stations and whose installed capacity and power generation reach a preset threshold are classified as main watersheds; the remaining sub-watersheds that are mainly composed of small and medium-sized distributed hydropower stations are classified as other watersheds.

[0011] As an alternative implementation, the adapted prediction model built for the main watershed is a dual-model fusion architecture; the dual-model fusion architecture includes: a scheduling simulation sub-model for simulating the hydrological cycle process of the watershed, and a deep learning sub-model for mining the nonlinear relationships of multi-source data; the output of the scheduling simulation sub-model is used as one of the inputs of the deep learning sub-model.

[0012] As an alternative implementation method, the scheduling simulation sub-model is an improved hydrological model that introduces a dynamic replenishment coefficient. The dynamic replenishment coefficient is dynamically adjusted based on the previous soil water storage and the current month's precipitation to optimize the calculation of surface water replenishment to soil water. The deep learning sub-model includes a multi-model parallel structure built based on multiple deep learning algorithms, and the prediction results of each model are fused through an ensemble learning strategy.

[0013] As an alternative implementation, the adapted prediction model built for other watersheds is a lightweight deep learning prediction model; the network structure complexity and feature input dimension of the lightweight deep learning prediction model are lower than those of the prediction model built for the main watersheds.

[0014] Secondly, the present invention provides a multi-physical factor coupled system for predicting the power generation of a cascade hydropower station in a river basin, comprising: The data acquisition and processing module is configured to: acquire multi-source data from the target area and preprocess it to build a high-quality dataset; The watershed division module is configured to divide the watersheds within the target area into different types based on the characteristics of hydropower distribution; The feature construction and filtering module is configured to: construct multi-physical factor feature vectors for different types of watersheds based on high-quality datasets, and perform feature filtering to obtain the optimal input feature set; The differentiated modeling and training module is configured to: build and train adaptive prediction models for different types of watersheds based on the optimal input feature set; The power generation prediction module is configured to input the multi-physical factor feature vector of the target time period into the trained and adapted prediction model, output and accumulate the predicted power generation values ​​of the sub-basin to obtain the total hydropower generation of the target area.

[0015] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0016] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0017] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a method for predicting the power generation of cascade hydropower stations in a river basin by coupling multiple physical factors. By introducing a dynamic replenishment coefficient to improve the ABCD hydrological model and coupling it with deep learning models such as LSTM, the physical mechanism and data-driven approach are deeply integrated. This method not only preserves the physical mechanism of the hydrological cycle but also leverages the advantages of deep learning in nonlinear mapping. It effectively solves the problems of missing physical mechanisms or instability caused by purely data-driven approaches in traditional methods, and achieves higher prediction stability under extreme weather conditions.

[0019] Furthermore, this invention, for the first time, divides watersheds into major watersheds and other watersheds for targeted modeling. A "dual-model fusion" architecture with high-dimensional features is designed for the major watersheds, while a "lightweight LSTM" model is designed for the other watersheds with sparse data. This differentiated modeling for each watershed improves prediction accuracy and solves the problem of homogeneity in traditional watershed modeling techniques. Verification through examples shows that the key indicators of this invention, such as R², MAPE, and RMSE, are significantly better than traditional models, meeting the requirements of high-precision engineering applications.

[0020] Furthermore, this invention uses a three-step method of "initial selection-optimization-dimensionality reduction" to screen features, combined with a self-designed feature importance score and an improved cross-correlation coefficient formula. This scientific feature screening effectively eliminates redundant features, avoids multicollinearity problems, and improves the training efficiency and generalization ability of the model.

[0021] Furthermore, the loss function of this invention is ingeniously designed and conforms to physical laws. It adopts a hybrid loss function with physical regularization, strengthens the learning of extreme samples through the dynamic weight MSE formula, and guides the prediction results to conform to physical conservation laws such as runoff-power generation and head-power generation through the coupling constraint formula, so that the prediction results are not only numerically accurate, but also more reasonable in physical meaning.

[0022] In summary, this invention automates the entire process from data acquisition, feature engineering, model training to prediction verification. It also includes a model update and optimization module that can dynamically adjust the model based on new data, ensuring the stability of prediction accuracy. It has strong engineering applicability, a high degree of automation, and fully meets the business needs of energy management departments to release forecast data on a monthly rolling basis. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method for predicting the power generation of a cascade hydropower station in a river basin by coupling multiple physical factors, as disclosed in an embodiment of the present invention. Figure 2 This is a comparison chart of the total predicted and observed power generation values ​​of a cascade hydropower station in a basin using the multi-physical factor coupling method of this invention, with basin A as the target area. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0025] The technical solutions disclosed in the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0026] Example 1 like Figures 1 to 2 As shown in the figure, this embodiment provides a method for predicting the power generation of a cascade hydropower station in a river basin by coupling multiple physical factors, including the following steps: Acquire multi-source data from the target region and preprocess it to construct a high-quality dataset; Based on the characteristics of hydropower distribution, the watersheds within the target area are divided into different types; Based on high-quality datasets, multi-physical factor feature vectors are constructed for different types of watersheds, and feature selection is performed to obtain the optimal input feature set; Based on the optimal input feature set, we construct and train adaptive prediction models for different types of watersheds. The multi-physical factor feature vectors of the target time period are input into the trained and adapted prediction model, and the predicted values ​​of power generation in each watershed are output and accumulated to obtain the total hydropower generation of the target area.

[0027] The specific solution of the present invention is as follows: This invention integrates multiple physical factors from meteorology, hydrology, geospatial data, and dispatching operations to implement differentiated modeling by watershed: Multi-source data is collected, preprocessed, and subjected to physical consistency screening to construct a high-quality dataset. The physical consistency screening uses clearly defined quantitative criteria, setting overall sequence Pearson correlation coefficient thresholds for core physical correlation pairs such as precipitation and runoff, reservoir level and outflow, and water consumption rate and power generation. The overall sequence positive correlation coefficient for precipitation and runoff is no less than 0.65, for reservoir level and outflow no less than 0.60, and for water consumption rate and power generation no less than 0.50. These serve as the basic criteria for determining compliance with overall physical laws. The total effective rainfall in the watershed is calculated using a grid-based installed capacity weighted method. Then, based on installed capacity, power generation ratio, and whether or not a controlling power station is included, the watershed is divided into major watersheds and other watersheds mainly composed of small and medium-sized distributed power stations. Features are screened through a three-step process of "initial selection, optimal selection, and dimensionality reduction," constructing high-dimensional and simplified feature vectors respectively. Subsequently, a dual-model fusion architecture of "scheduling simulation + deep learning" was built for the main river basins, and a lightweight LSTM model was designed for other river basins. The sample set was then divided into a 7:1.5:1.5 ratio, and the model was trained using a hybrid loss function with physical regularization, coupled with anti-overfitting and early stopping mechanisms. Finally, the feature vectors were input to complete the sub-basin prediction and accumulate the power generation. The accuracy was verified using multiple indicators such as R², MAPE, and RMSE; if the standards were not met, adjustments were made retrospectively. This invention is applicable to the prediction of total hydropower generation in all river basins nationwide for the next 1-6 months, and can support scenarios such as energy dispatch optimization. It solves the problems of existing technologies such as the disconnect between physical mechanisms and data-driven approaches, homogenized modeling, and insufficient engineering applicability.

[0028] (I) Multi-source data acquisition and high-quality dataset construction. Meteorological data, hydrological data, hydropower station operation data, and geospatial data of the target area are collected. Through data preprocessing and physical consistency screening, outlier data is removed and missing values ​​are filled to construct a high-quality dataset that meets the requirements of model training. The multi-source data covers four categories: First, meteorological data, including historical monitoring data and medium- and long-term forecast data such as rainfall, temperature, relative humidity, and wind speed; second, hydrological data, including watershed hydrological element data such as inflow, inter-regional inflow, and soil moisture content; third, hydropower station operation data, including operation indicators such as installed capacity, monthly power generation, reservoir water level, outflow, and water consumption rate; and fourth, geospatial data, including spatial basic data such as digital elevation model (DEM), water system GIS layer, and catchment area boundaries.

[0029] The rainfall data used in this embodiment is the effective rainfall data. The effective rainfall data is calculated using a self-designed grid installed capacity weighted method, and the core formula is as follows: 1. Calculation of grid installed capacity weighting coefficient: ; in, For the first The weight coefficients of each grid, For the first The set of reservoir stations corresponding to all catchment areas of each grid. For reservoir stations The installed capacity, This refers to the set of all reservoir stations within the target area.

[0030] 2. Calculation of grid terrain correction coefficient: ; in For the first The average slope of each grid, For the first Average elevation (m) of each grid.

[0031] 3. Calculation of effective rainfall in the grid: ; in For the first Effective rainfall per grid For the first The measured rainfall in each grid.

[0032] 4. Calculation of total effective rainfall in the watershed: ; in This represents the total effective rainfall in the basin. A set of grids within the watershed area. For the first The area of ​​each grid, This represents the total area of ​​the basin.

[0033] (II) Basin Classification and Multi-Physical Factor Feature Engineering. Based on the distribution characteristics and power generation ratio of hydropower bases in the target area, the basins are divided into two categories: major basins, which include control cascade power stations and account for no less than 60% of power generation; and other basins, which are mainly composed of small and medium-sized distributed hydropower stations. To address the differences in hydrological characteristics and dispatching capabilities between the two types of basins, multi-physical factor feature vectors are constructed for each category. Feature selection is completed through a three-step process of "initial selection - optimal selection - dimensionality reduction" to obtain the optimal input feature set. In terms of feature vector construction, high-dimensional feature vectors are constructed for major basins, covering meteorological features, hydrological features, dispatching features, and power station characteristics, comprehensively reflecting the basin's hydrological cycle and dispatching control patterns. Concise feature vectors are constructed for other basins, focusing on meteorological features, basic power station characteristics, and historical power generation characteristics, adapting to the limited dimensionality of distributed power station data. In the feature selection process, the initial selection uses Pearson correlation analysis to retain features with a correlation coefficient of at least 0.3 with power generation; the next selection uses the random forest algorithm to calculate feature importance, retaining the top 80% of core features. The feature importance score uses a self-designed formula. ; in, Features Importance rating For the number of decision trees, For the first The out-of-bag error of a decision tree For the first Out-of-bag random permutation features of decision trees The error after that, The standard deviation of the out-of-bag data error for all decision trees is used. Based on the score, the top 80% of the highly important features are retained to reduce the computational burden caused by redundant features. At the same time, core features that are strongly correlated with power generation prediction are retained to provide high-quality input for subsequent model training and avoid irrelevant features from interfering with prediction accuracy.

[0034] The core method for calculating feature importance using conventional random forests is the average reduction of out-of-bag errors. This method has two significant drawbacks, particularly prominent in multi-physical factor feature engineering for watershed hydropower: 1) Lack of standardization, resulting in incomparable scores and poor robustness: The absolute value of out-of-bag errors is greatly affected by dataset size, feature dimension, and number of decision trees. Scores for different features can only be compared relatively within the same calculation and cannot be reused across datasets / feature sets. Furthermore, when the overall out-of-bag error fluctuates significantly (e.g., in hydropower prediction where data from small and medium-sized watersheds is sparse and sample noise is high), it easily amplifies small error changes, leading to secondary features being misclassified as core features, thus diluting the scores of core features. 2) Failure to consider the overall error distribution, making it sensitive to extreme errors: Conventional formulas directly accumulate error differences. If some decision trees exhibit extreme out-of-bag errors due to sample noise, it directly affects the final result of feature importance. In scenarios with temporal fluctuations in hydropower meteorological and hydrological features, this can easily lead to misjudgments of feature importance. The self-designed feature importance scoring formula adds a standardization term to the conventional formula. The key improvement is to sum the error differences and divide by the product of the number of decision trees and the standard deviation of the out-of-bag data error. The score is then globally standardized using std(Errt), and the number of decision trees is normalized by combining T.

[0035] Dimensionality reduction employs cross-correlation analysis, and the cross-correlation coefficient uses an improved sliding window formula: ; in, Features and Lag The cross-correlation coefficient of the steps, For covariance, For variance, This represents the time step. The closer the coefficient is to 1 or -1, the stronger the linear correlation between the two features; when... When the correlation coefficient is ≥0.95, it is determined that there is serious redundancy between the two features, and one of the features is removed to avoid multicollinearity between features, reduce the complexity of model training, and at the same time ensure the independence and effectiveness of the feature set, thereby improving the generalization ability and training efficiency of the subsequent prediction model. The conventional cross-correlation coefficient formula is a correlation analysis method for static data, without a time lag term, and has obvious adaptability defects in the dimensionality reduction of multiple physical factor features of watershed hydropower time series. The self-designed sliding window improved cross-correlation coefficient formula adds a sliding lag step τ, which can flexibly adjust the window range according to the time scale of watershed hydropower prediction (such as monthly scale, daily scale); it maintains the temporal nature of the features and analyzes the correlation between two features at different time steps, rather than only at the same time step.

[0036] After completing the feature selection through the above three-step method of "initial selection-optimization-dimensionality reduction", the optimal input feature set is finally determined by minimizing the RMSE on the validation set through recursive feature elimination (RFE) to make the feature combination that makes the model have the strongest generalization ability.

[0037] (III) Construction of Differentiated Prediction Models. Adaptive prediction model architectures were designed for the characteristics of the two types of watersheds, achieving a deep integration of physical mechanisms and deep learning algorithms. The main watersheds adopted a dual-model fusion architecture of "scheduling simulation + deep learning": the core of the scheduling simulation sub-model is an improved three-layer ABCD hydrological model, constructing a three-layer monthly water balance system of surface water, soil water, and groundwater. A self-designed dynamic replenishment coefficient was introduced to optimize model accuracy, replacing the traditional fixed replenishment coefficient. The formula for the dynamic replenishment coefficient is: ; in, For the first Monthly surface water recharge coefficient to soil water For the first Monthly soil water storage This represents the maximum soil moisture content. For the first Monthly precipitation, This represents the average monthly precipitation over many years.

[0038] The amount of surface water introduced to replenish soil water is calculated as follows: ; in, This refers to the amount of surface water replenishing soil water. For the first Monthly surface water storage, For the first monthly evapotranspiration. After the calculation is completed, it is substituted into the soil water balance system of the improved ABCD model to participate in... The update will further optimize the inflow runoff optimization results.

[0039] Combining LSTM neural network correction of runoff prediction results, the correction model adopts a self-designed error feedback mechanism expression: ; in, For the corrected first Monthly runoff forecast, To improve the predictions of the ABCD model, This is the error in the previous month's runoff forecast. This represents the anomaly of precipitation in that month.

[0040] The deep learning sub-model is constructed based on three classic algorithms: LSTM, ConvLSTM, and XGBoost, forming a multi-model parallel structure. It combines Bayesian optimization to optimize model hyperparameters and introduces a self-designed attention mechanism to enhance the weights of key features. The attention mechanism formula is as follows: ; ; in, For the input feature matrix, This is the normalization function used for the attention mechanism. , This is the weight matrix. , For bias terms, For feature dimension, This is the feature matrix after attention weighting.

[0041] After modeling using LSTM, ConvLSTM, and XGBoost respectively, the prediction results of each model are fused through a stacking ensemble learning strategy. The meta-classifier uses logistic regression, and the ensemble prediction weights adopt a self-designed error adaptive formula. ; The ensemble prediction results are as follows: ; in, To integrate the prediction results, (m=1, 2, 3...) represents the nth The weights of each model, , , These are the predicted values ​​for the LSTM, ConvLSTM, and XGBoost models, respectively. For the first The mean absolute percentage error of the model was determined. This attention mechanism initializes the attention weights based on prior knowledge from hydropower engineering experts. Initial attention weights for hydrological, meteorological, and physical factors that play a crucial role in power generation, such as rainfall, inflow, and reservoir water level, are assigned high weights (0.7–0.9), while secondary features are assigned low weights (0.1–0.3). Through model training and experimental verification, the mechanism adaptively strengthens the attention weights for key physical factors such as rainfall, inflow, and soil water storage during iteration. Ultimately, the core physical factors emphasized align closely with the experience and judgments of experts in hydropower dispatching, achieving a creative application of the attention mechanism in the specific scenario of power generation prediction for cascade hydropower stations in a river basin.

[0042] (iv) Optimization of Model Training Loss Function. The optimal feature set is divided into training, validation, and test sets in a 7:1.5:1.5 ratio. A hybrid loss function with physical regularization is used to train the model. An early stopping mechanism and an anti-overfitting module are also introduced to ensure the model's generalization ability. The hybrid loss function consists of three parts, with each part's contribution adjusted by weight coefficients. The data fitting loss uses a self-designed dynamic weight formula: ; in, For data fitting loss, For sample size, To predict power generation, This represents the actual amount of electricity generated. This represents the average actual power generation. The standard deviation of actual power generation is used to dynamically adjust the loss weights for different power generation levels using an exponential function; the physical constraint regularization term adopts a self-designed coupled constraint formula. ; in, To predict runoff, To predict water head, , , These are historical average runoff, average power generation, and average head, respectively. , This represents the physical constraint coefficient.

[0043] Based on the above, the formula for the hybrid loss function with physical regularization is: ; in For the total loss function, This represents the physical constraint coefficient.

[0044] (V) Power Generation Prediction and Accuracy Verification. The multi-physical factor feature vectors for the target time period are input into the trained differentiated model, which outputs predicted power generation values ​​for the main basin and other basins respectively. These are then summed to obtain the total hydropower generation for the target area. A multi-index system is used to comprehensively verify the prediction accuracy. The core compliance indicators are: a test set determination coefficient (R²) of no less than 0.82, a mean absolute percentage error (MAPE) of no more than 18%, and a root mean square error (RMSE) of no more than 1.2 billion kWh, ensuring that the prediction results meet the requirements for engineering applications. The system has a built-in automated backtracking optimization mechanism that requires no manual intervention. If the prediction accuracy does not meet the aforementioned core performance indicators, the model optimization process will be automatically initiated: First, a Bayesian optimization algorithm will be used to iteratively search within the preset hyperparameter space to optimize the model's hyperparameters and retrain the model. If the prediction accuracy still does not meet the requirements after hyperparameter optimization, the feature selection process will be re-executed, expanding the candidate feature range from the original optimal input feature set to 90% of the feature set after initial selection. Potential core features will be added, and the feature vector will be reconstructed and the model training will be completed. If the accuracy still does not meet the requirements after hyperparameter optimization and feature re-selection, the system will issue an early warning and support technical personnel to intervene for manual tuning and parameter correction until the prediction accuracy meets the core performance indicators required for engineering applications.

[0045] The technical solution of the present invention will be described in detail below with reference to specific examples. This example takes the A basin as the target area. The area is a mountainous canyon terrain in the southwest, with a total basin area of ​​approximately 473,000 km². It includes the B, C, and D cascade hydropower stations, with a total installed capacity of over 70 million kW. The power generation of the cascade hydropower stations accounts for 75% of the total power generation of the basin. It is the core hydropower base in southwest my country. Those skilled in the art can adjust the parameters according to the actual scenario and are not limited to the specific settings of this embodiment.

[0046] (I) Multi-source data acquisition and high-quality dataset construction. Regarding data sources, meteorological data utilizes monitoring data from surface meteorological stations of the China Meteorological Administration (2003-2022, monthly scale) and ECMWF medium- and long-term forecast data (spatial resolution 0.5°×0.5°), covering four categories of indicators: precipitation (monthly average 0~890mm), temperature (monthly average -5℃~28℃), relative humidity (monthly average 45%~95%), and wind speed (monthly average 0.8~6.2m / s). Hydrological data originates from the hydrological station monitoring network of the Water Resources Commission, including inflow runoff (monthly average 1200~18500m³ / s) and inter-regional inflow (monthly average 200~3500m³ / s). Soil moisture content (monthly average 15%~45%); Hydropower station operation data are obtained from the power station monitoring system, covering installed capacity (3 million~16 million kW per station), monthly power generation (0~8.5 billion kWh per station), reservoir water level (normal storage level 825~1200m), outflow (monthly average 800~15000 m³ / s), and water consumption rate (3.2~5.8 m³ / kW·h); Geospatial data uses 30m resolution DEM data (sourced from the Geospatial Data Cloud), 1:250,000 water system GIS layer, and catchment area boundary vector data (sourced from the National Geographic Information Public Service Platform). Data preprocessing employed the 3σ criterion to remove outliers, resulting in the removal of 326 meteorological anomalies, 189 hydrological anomalies, and 97 operational anomalies. Missing value imputation used linear interpolation to fill 582 short-term (1-7 day) missing values, and random forest interpolation to fill 215 long-term (8-30 day) missing values. Physical consistency screening verified a positive correlation between precipitation and runoff (Pearson correlation coefficient 0.82), a match between reservoir water level and outflow (correlation coefficient 0.78), and a negative correlation between water consumption rate and power generation (correlation coefficient -0.65), eliminating 103 samples that did not conform to physical laws. The total effective rainfall of the watershed was calculated using the grid-based installed capacity weighted method. The watershed was divided into 1.892 million 500m × 500m grids. The grid installed capacity weight coefficient ranged from 0.001 to 0.085, the grid topography correction coefficient ranged from 1.02 to 1.89 (average slope 15° to 45°, average elevation 1200 to 4500m), and the grid effective rainfall ranged from 0 to 925mm. Finally, the monthly average total effective rainfall of watershed A was calculated to be 5 to 865mm.

[0047] (II) Basin Classification and Multi-Physical Factor Characterization. Based on the distribution characteristics of hydropower bases in the basin, Basin A is divided into the main basin (the area covered by the BCD cascade hydropower stations, accounting for 75% of the total installed capacity and 78% of the power generation, including key cascade hydropower stations) and other basins (the area of ​​small and medium-sized distributed hydropower stations on tributary A, with each station having an installed capacity of less than 50,000 kW, totaling 213 stations). A high-dimensional feature vector with 68 indicators is constructed for the main basin, and a simplified feature vector with 22 indicators is constructed for the other basins. Feature selection employs a three-step method: initial selection, optimal selection, and dimensionality reduction. Initial selection uses Pearson correlation analysis to retain features with a correlation coefficient ≥0.3 with power generation, retaining 52 dimensions for major river basins and 18 dimensions for other river basins. Optimization uses a random forest algorithm (number of decision trees T=500) combined with a self-designed feature importance scoring formula to retain the top 80% of core features, retaining 41 dimensions for major river basins and 14 dimensions for other river basins, with feature importance scores ranging from 0.02 to 0.95. Dimensionality reduction uses a sliding window-based improved cross-correlation analysis (sliding window step size τ=3 months) to eliminate redundant features with an absolute cross-correlation coefficient ≥0.95, ultimately obtaining a 32-dimensional optimal input feature set for major river basins and a 10-dimensional optimal input feature set for other river basins, without feature multicollinearity issues.

[0048] (III) Construction of Differentiated Prediction Models. A dual-model fusion architecture of "scheduling simulation + deep learning" is constructed for major watersheds: the scheduling simulation sub-model is an improved three-layer ABCD hydrological model, which introduces a dynamic replenishment coefficient. The calculated value of this coefficient is 0.3~0.98 (soil water storage). The value ranges from 120 to 480 mm, representing the maximum soil moisture content. =500mm, monthly precipitation Values ​​range from 0 to 890 mm, representing the average monthly precipitation over the past 30 years. =125mm), surface water recharge to soil water Value range: 5~780mm, evaporation rate The monthly average runoff was calculated using the Penman-Monteith formula to be 8-65 mm. Runoff correction employed an LSTM neural network with a 3-dimensional input layer, 3 hidden layers with 128, 64, and 32 neurons respectively, and a 1-dimensional output layer. The learning rate was 0.001, and the batch size was 32. The error feedback mechanism incorporated the previous month's runoff prediction error. The value ranges from -2500 to 3200 m³ / s, representing the monthly precipitation anomaly. The value range is -120~765mm. After correction, the correlation coefficient between the predicted and measured runoff values ​​improved from 0.75 to 0.89. A multi-model parallel structure of LSTM, ConvLSTM, and XGBoost was constructed using deep learning sub-models. The hyperparameters were optimized using a Bayesian optimization algorithm (100 iterations, learning rate 0.0005~0.002, tree depth 3~8). The weight matrix in the self-designed attention mechanism... Dimensions 32×64 Dimensions 64×32, bias term , Value range: -0.5 to 0.5, feature dimension =32, and after attention weighting, the weights of key features (precipitation, inflow runoff) are increased to 0.85~0.92. Stacking ensemble learning calculates the MAPE of each model based on the validation set: LSTM model MAPE=12.5%, ConvLSTM model MAPE=10.8%, XGBoost model MAPE=11.2%, and the weights are calculated using the error adaptive weighting formula. =0.31、 =0.36、 =0.33, satisfying the weight sum of 1. A lightweight LSTM model is constructed for other watersheds, with a 10-dimensional simplified feature vector as the input layer and two hidden layers, with 64 and 32 neurons respectively. The attention mechanism and multi-model parallel structure are omitted. The Adam optimizer (learning rate 0.001) is used for training, with a batch size of 16, which is suitable for the characteristics of small data volume of distributed power stations (the sample size is only 30% of that of the main watersheds).

[0049] (iv) Optimization of the model training loss function. The optimal feature set is divided into training, validation, and test sets in a 7:1.5:1.5 ratio. All feature data are mapped to the [0,1] interval using Min-Max normalization. A hybrid loss function with physical regularization is used to set physical constraint coefficients. =0.3, in the dynamic weights MSE formula for data fitting loss, the sample size is... =240, average actual power generation =11 billion kWh, standard deviation of actual power generation =5.5 billion kWh, with dynamic weights ranging from 0.05 to 0.98. Samples with extreme power generation (monthly power generation > 20 billion kWh or < 5 billion kWh) are assigned higher weights (above 0.85). The physical constraint regularization term sets sub-constraint coefficients α = 0.7 and β = 0.3, based on historical average runoff. =8500 m³ / s, historical average power generation =11.2 billion kWh, historical average head =185m, the coupling error between predicted runoff and power generation is ≤8%, and the coupling error between predicted head and power generation is ≤5%. An anti-overfitting module incorporates a Dropout layer (Dropout probability = 0.3), with an L2 regularization coefficient of 5e-4. The early stopping mechanism is set to a patience value of 8 training epochs; training stops when the validation set loss decreases by <0.001 for 8 consecutive epochs. The model ultimately reaches its optimum in the 62nd epoch, with the validation set loss stabilizing at 0.032.

[0050] like Figure 2 This is a comparison chart of the predicted and observed power generation values ​​for the entire basin, using the multi-physical factor coupling method for predicting power generation in a basin-wide cascade hydropower station, as described in this invention, with basin A as the target area. The input data consists of high-dimensional feature vector data for 240 months from 2003 to 2022 for basin A. Figure 2 The horizontal axis represents the number of months, with 0 corresponding to January 2003 and 240 corresponding to December 2022; the vertical axis represents the model's total power generation prediction and observed power generation, where the model's total power generation prediction is the sum of the monthly power generation model predictions for the main river basins and the monthly power generation model predictions for other river basins.

[0051] (V) Power Generation Forecast and Accuracy Verification. The multi-physical factor feature vectors for January-June 2025 were input into the trained differential model, outputting monthly power generation forecasts of 4.5-8.2 billion kWh for major river basins and 800-1.8 billion kWh for other river basins. These forecasts were summed to obtain a total monthly power generation forecast of 5.3-10 billion kWh for River Basin A. The accuracy of the test set was verified using a multi-index system. The calculated coefficient of determination R² = 0.89, mean absolute percentage error (MAPE) = 12.5%, and root mean square error (RMSE) = 860 million kWh, all meeting the accuracy requirements of this invention: R² ≥ 0.82, MAPE ≤ 18%, and RMSE ≤ 1.2 billion kWh. Compared with traditional hydrological models (improved ABCD model) and single LSTM models, the R2 of the model in this invention is improved by 0.18, MAPE is reduced by 9.2%, and RMSE is reduced by 630 million kWh, significantly improving prediction accuracy. It can effectively support the medium- and long-term prediction of power generation of cascade hydropower stations in the A basin for 1 to 6 months and the optimization of energy dispatch.

[0052] Example 2 This embodiment provides a multi-physical factor coupled system for predicting the power generation of a cascade hydropower station in a river basin, including: The data acquisition and processing module is configured to: acquire multi-source data from the target area and preprocess it to build a high-quality dataset; The watershed division module is configured to divide the watersheds within the target area into different types based on the characteristics of hydropower distribution; The feature construction and filtering module is configured to: construct multi-physical factor feature vectors for different types of watersheds based on high-quality datasets, and perform feature filtering to obtain the optimal input feature set; The differentiated modeling and training module is configured to: build and train adaptive prediction models for different types of watersheds based on the optimal input feature set; The power generation prediction module is configured to input the multi-physical factor feature vector of the target time period into the trained and adapted prediction model, output and accumulate the predicted power generation values ​​of the sub-basin to obtain the total hydropower generation of the target area.

[0053] It should be noted that the above modules correspond to the steps in Embodiment 1, and the examples and application scenarios implemented by the above modules and their corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules can be executed in a computer system as part of the system.

[0054] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0055] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0056] A computer-readable storage medium for storing computer instructions that, when executed by a processor, perform the method of Embodiment 1.

[0057] The method in Example 1 can be directly executed by a hardware processor, or it can be executed by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0058] A computer program product includes a computer program that, when executed by a processor, implements the method in Embodiment 1.

[0059] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0060] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0061] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0062] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein 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.

[0063] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for predicting the power generation of cascade hydropower stations in a river basin based on the coupling of multiple physical factors, characterized in that, Includes the following steps: Acquire multi-source data from the target region and preprocess it to construct a high-quality dataset; Based on the characteristics of hydropower distribution, the watersheds within the target area are divided into different types; Based on the high-quality dataset, multi-physical factor feature vectors are constructed for different types of watersheds, and feature filtering is performed to obtain the optimal input feature set; Based on the optimal input feature set, adaptive prediction models are constructed and trained for different types of watersheds; The multi-physical factor feature vectors of the target time period are input into the trained and adapted prediction model, and the predicted values ​​of power generation in each watershed are output and accumulated to obtain the total hydropower generation of the target area.

2. The method for predicting the power generation of a cascade hydropower station in a river basin based on the coupling of multiple physical factors as described in claim 1, characterized in that, The construction of the high-quality dataset further includes calculating the total effective rainfall in the watershed using a grid-based installed capacity weighting method. The grid-based installed capacity weighting method includes: calculating the grid installed capacity weighting coefficient based on the installed capacity of hydropower stations within the grid; calculating the grid terrain correction coefficient based on the grid terrain features; calculating the grid effective rainfall by combining the measured rainfall, grid installed capacity weighting coefficient, and grid terrain correction coefficient; and finally, obtaining the total effective rainfall of the watershed by weighting and summing the grid areas.

3. The method for predicting the power generation of a cascade hydropower station in a river basin based on the coupling of multiple physical factors as described in claim 1, characterized in that, The specific method of dividing the watersheds within the target area into different types is as follows: sub-watersheds containing control cascade hydropower stations and whose installed capacity and power generation reach a preset threshold are classified as main watersheds; the remaining sub-watersheds, mainly composed of small and medium-sized distributed hydropower stations, are classified as other watersheds.

4. The method for predicting the power generation of a cascade hydropower station in a river basin based on the coupling of multiple physical factors as described in claim 3, characterized in that, The adapted prediction model constructed for the main watershed is a dual-model fusion architecture; The dual-model fusion architecture includes: a scheduling simulation sub-model for simulating the hydrological cycle process of a watershed, and a deep learning sub-model for mining nonlinear relationships between multi-source data. The output of the scheduling simulation sub-model is used as one of the inputs to the deep learning sub-model.

5. The method for predicting the power generation of a cascade hydropower station in a river basin based on the coupling of multiple physical factors as described in claim 4, characterized in that, The scheduling simulation sub-model is an improved hydrological model that introduces a dynamic replenishment coefficient. The dynamic replenishment coefficient is dynamically adjusted based on the previous soil water storage and the current month's precipitation to optimize the calculation of surface water replenishment to soil water. The deep learning sub-model includes a multi-model parallel structure built based on multiple deep learning algorithms, and the prediction results of each model are fused through an ensemble learning strategy.

6. The method for predicting the power generation of a cascade hydropower station in a river basin based on the coupling of multiple physical factors as described in claim 3, characterized in that, The adapted prediction model built for other watersheds is a lightweight deep learning prediction model; the network structure complexity and feature input dimension of the lightweight deep learning prediction model are lower than those of the prediction model built for the main watersheds.

7. A watershed cascade hydropower station power generation prediction system with multi-physical factor coupling, characterized in that, include: The data acquisition and processing module is configured to: acquire multi-source data from the target area and preprocess it to build a high-quality dataset; The watershed division module is configured to divide the watersheds within the target area into different types based on the characteristics of hydropower distribution; The feature construction and filtering module is configured to: construct multi-physical factor feature vectors for different types of watersheds based on the high-quality dataset, and perform feature filtering to obtain the optimal input feature set; The differentiated modeling and training module is configured to: construct and train suitable prediction models for different types of watersheds based on the optimal input feature set; The power generation prediction module is configured to input the multi-physical factor feature vector of the target time period into the trained and adapted prediction model, output and accumulate the predicted power generation values ​​of the sub-basin to obtain the total hydropower generation of the target area.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-6.

Images