Basin hydroelectric power generation prediction method based on multi-source data fusion and deep learning

Abstract

The application relates to the technical field of water and electricity energy prediction, and discloses a watershed water and electricity power generation prediction method based on multi-source data fusion and deep learning, which comprises the following steps: collecting data information of water and electricity stations in a watershed, performing data fusion and space-time alignment on the data information, and constructing a meteorology-hydrology-engineering coupled data set; a grid scale dynamic effective rainfall coefficient model is constructed, and an effective rainfall space-time sequence considering space-time non-stationarity is calculated; according to the meteorology-hydrology-engineering coupled data set and the effective rainfall space-time sequence, multi-scale space-time feature engineering is performed; a multi-task physical constraint deep learning model with a shared feature extraction layer is established to predict water and electricity power generation; and the hyperparameters of the multi-task physical constraint deep learning model are optimized and updated online. By using the application scheme, the prediction accuracy and the model robustness are significantly improved, and reliable decision support is provided for complex watershed water and electricity dispatching.

Description

Technical Field

[0001] This invention relates to the field of hydropower energy prediction technology, specifically to a method for predicting watershed hydropower generation based on multi-source data fusion and deep learning. Background Technology

[0002] Hydropower, as a clean and renewable energy source, plays a crucial role in peak shaving, frequency regulation, and capacity support in new power systems. Accurate forecasting of hydropower generation is essential for grid optimization, electricity market trading, and extreme weather response. Traditional forecasting methods mainly include: statistical regression methods (such as ARIMA models and multiple linear regression), which struggle to capture nonlinear spatiotemporal dependencies; physical hydrological models (such as SWAT models and VIC models), which are complex in structure, have numerous parameters, require high data quality, and are computationally time-consuming; and single deep learning methods (such as standard LSTM and GRU), which do not fully consider the spatial heterogeneity of watersheds and engineering constraints, and lack physical interpretability. Furthermore, these existing technologies have the following shortcomings:

[0003] Spatial coupling effect is missing: the watershed is treated as a homogeneous body, and the spatial overlap and water competition between multiple reservoir clusters are ignored.

[0004] Static engineering status: The impact of dynamic changes in engineering status such as reservoir storage and flood control level on rainfall-to-power conversion efficiency is not considered;

[0005] The physical mechanisms are black boxed: Purely data-driven models cannot guarantee that the prediction results conform to basic physical laws such as water balance and energy conversion;

[0006] Weak online adaptability: The model parameters are fixed and it is difficult to adapt to climate change and underlying surface evolution.

[0007] To address this, we invented a method for predicting hydropower generation in river basins based on multi-source data fusion and deep learning, which solves the above-mentioned technical problems. Summary of the Invention

[0008] This invention provides a method for predicting hydropower generation in river basins based on multi-source data fusion and deep learning. By constructing a complete "data-model-application" technology system, it significantly improves the prediction accuracy, robustness, and interpretability of hydropower generation in complex river basins.

[0009] Therefore, the present invention provides the following technical solution:

[0010] A method for predicting hydropower generation in a river basin based on multi-source data fusion and deep learning, the method comprising:

[0011] Step 1: Collect data on hydropower stations within the basin, and perform data fusion and spatiotemporal alignment to construct a coupled meteorological-hydrological-engineering dataset;

[0012] Step 2: Construct a grid-scale dynamic effective rainfall coefficient model and calculate the spatiotemporal sequence of effective rainfall considering spatiotemporal nonstationarity;

[0013] Step 3: Perform multi-scale spatiotemporal feature engineering based on the meteorological-hydrological-engineering coupled dataset and the effective rainfall spatiotemporal sequence;

[0014] Step 4: Establish a multi-task physical constraint deep learning model with a shared feature extraction layer to predict hydropower generation.

[0015] Step 5: Optimize and update the hyperparameters of the multi-task physical constraint deep learning model online.

[0016] Optionally, in step 1, the collected data on hydropower stations within the basin includes meteorological observation data, hydrological monitoring data, engineering operation data, and remote sensing inversion data. The meteorological observation data includes hourly rainfall, temperature, humidity, wind speed, and radiation; the hydrological monitoring data includes daily soil moisture, daily snowmelt water equivalent, and daily evaporation; the engineering operation data includes hourly reservoir water level, inflow, outflow, actual power generation, installed capacity, flood control limit water level, and design reservoir capacity; and the remote sensing inversion data includes GPM satellite precipitation products and MODIS surface temperature and evapotranspiration products. When fusing and aligning the data in time and space, a triple-nested grid system is used for spatial downscaling, and inverse distance weighted interpolation and terrain correction are used to achieve data assimilation from stations to the grid. In terms of time scale, wavelet packet decomposition and reconstruction are used to achieve unified temporal granularity alignment of data at different frequencies, constructing a meteorological-hydrological-engineering coupled dataset with unified spatiotemporal resolution.

[0017] Optionally, in step 2, based on the spatial topological relationship and installed capacity weight of each hydropower station's catchment area, a time decay factor and a reservoir storage state correction factor are introduced to construct the grid-scale dynamic effective rainfall coefficient model, and the spatiotemporal sequence of effective rainfall considering spatiotemporal nonstationarity is calculated; wherein, the grid-scale dynamic effective rainfall calculation model considering the dynamic storage state of the reservoir group is as follows:

[0018]

[0019] in:

[0020] For grid ( x , y At that moment t Effective rainfall;

[0021] Original rainfall data from the grid;

[0022] The spatial effective rainfall coefficient is calculated using the following formula:

[0023]

[0024]

[0025] In the formula, C i For the first i The installed capacity of each hydropower station; As an indicator function, whether the grid is in the first position i Within the catchment area of ​​a power station, if it is, the value is 1; otherwise, the value is 0. For the first i Power station t The availability of storage capacity at any given time. For the first i Each power station at any time t The actual storage capacity For the first i Total designed reservoir capacity of the power station;

[0026] β ( t The time decay factor is used, and the Gamma distribution function is employed to characterize the lag effect of rainfall-runoff.

[0027] This is a correction factor for the energy storage status, reflecting the amplification / inhibition effect of the overall regulation and storage capacity of the reservoir group on effective rainfall.

[0028] Optionally, in step 3, multi-scale wavelet decomposition is performed on the effective rainfall, daily soil moisture, daily snowmelt water equivalent, daily evaporation, water level fluctuation, and inflow, extracting the features of trend, periodic, and residual terms, and constructing a spatiotemporal lag cross factor, specifically including:

[0029] Step 3.1: Perform three-level wavelet decomposition on the effective rainfall spatiotemporal sequence to obtain low-frequency trend component A3 and high-frequency detail components D1, D2, and D3, which respectively represent long-term trend, seasonal cycle, and random disturbance. After removing the random disturbance, the components are synthesized to obtain the effective rainfall sequence with random interference eliminated.

[0030] Step S3.2: Construct the spatiotemporal lag crossover factor matrix F lag Considering the nonlinear lag relationship between rainfall and power generation:

[0031]

[0032]

[0033]

[0034]

[0035] In the formula, This represents the total effective rainfall in the basin. X × Y This represents the total number of grid cells in the watershed. This refers to the total inflow into the reservoir, which is the sum of the inflows from all hydropower stations. ( t ) is the first i Power station t Real-time inflow rate (unit: m³ / s); The average hydropower head for power generation in the basin is calculated by weighting the installed capacity. For the first i The installed capacity of each power station; For the first i Power station t Constantly generating water head; L R for The lag order is determined by the maximum mutual information coefficient (MIC), with a value ranging from 0 to 12 months. L Q The lag order of the flow is determined by the maximum mutual information coefficient (MIC), with a value ranging from 0 to 6 months. L H The lag order of the water head is determined by the maximum mutual information coefficient MIC, with a value ranging from 0 to 3 months.

[0036] Every moment t of For (1+ L R + L Q + L H A 3D vector, the feature matrix constructed over all time series. ;

[0037] Step 3.3: Construct spatial interaction features and calculate the spatial convolution features of effective rainfall in the grid and available storage capacity of the upstream reservoir:

[0038]

[0039]

[0040]

[0041] In the formula, No. i The first power station tThe spatial interaction feature vector at any given time represents the combined impact of upstream water inflow and storage capacity on this power station; i Index for the target power station, i =1,2,..., N ; j For grids or power plant indexes within the upstream influence domain; For the first i The upstream influence domain of a power station is defined as the directly connected upstream power station. j ; For spatial influence weighting, a combined calculation of flow direction and distance attenuation is used; For flow direction indication function, if the power station j Water flows towards the power station i but =1, otherwise =1 =0; For power station j to the power station i The length of the water flow path, in km; D 0 represents the attenuation characteristic length, which is taken as 3 times the average reservoir spacing in the basin; For upstream power stations j At any moment t Available storage capacity; For power station j Total storage capacity; This refers to the availability of storage capacity.

[0042] Step 3.4: Construct a unified spatiotemporal input tensor, and convert the feature matrix constructed on the watershed-level time series. Interactive feature vectors of power plant-level space By splicing the data, a three-dimensional input tensor is constructed. Among them, the time dimension T The value is set to 24 months to include at least two complete wet and dry water cycles and ensure sufficient preceding information; spatial dimension N The number of power stations in the basin; characteristic dimension D =2+ L R + L Q + L H .

[0043] Optionally, in step 4, a multi-task physical constraint deep learning model with a shared feature extraction layer is established. The main task predicts hydropower generation, and the auxiliary task predicts the available water volume in the basin. A physical consistency constraint term based on the water balance equation and the power generation function is embedded in the loss function of the main task. The multi-task physical constraint deep learning model is an end-to-end single-task deep learning architecture, and its input data is the three-dimensional input tensor. Output the current time.t The predicted total power generation of the basin .

[0044] Optionally, in step 4, the multi-task physical constraint deep learning model includes an input layer, a feature extraction layer, and a prediction output layer;

[0045] The input layer receives the three-dimensional input tensor X, which is constructed following the feature availability principle of the prediction stage, specifically including: the current time step. t Total effective rainfall in the basin Known data for historical moments: t- 1, t- 2,…, tL R Effective rainfall at any time t- 1,…, tL Q Inbound traffic at any time t- 1,…, tL H The constant head of the generator;

[0046] The feature extraction layer uses a 3-layer ConvLSTM to capture a space-time coupled pattern, with each layer followed by a spatiotemporal attention mechanism. The attention weights are calculated as follows:

[0047]

[0048] Where σ is the sigmoid activation function. Here is the attention weight matrix. For bias vectors, This is the hidden state from the previous moment. For the current input features;

[0049] The prediction output layer uses a two-layer fully connected network, with the final hidden state of ConvLSTM. As input, where Output the number of hidden nodes in the ConvLSTM layer and the current time step. t The predicted total power generation of the basin ;

[0050] The main task loss function is constructed, consisting of two parts: deep learning error and physical consistency constraint.

[0051]

[0052]

[0053]

[0054] In the formula, LThe main task loss function; The mean square error of power generation prediction; The physical consistency constraint term is constructed based on the fundamental equations of hydropower conversion. for t Actual observed values ​​of power generation at any given time. T This represents the total length of the time series. These are the physical constraint weighting coefficients; η is the overall efficiency coefficient of the hydropower station, which is dimensionless; g is the acceleration due to gravity. for t The average power generation flow of the basin at any given time is calculated using measured data. for t The average hydropower head of the basin at any given time is calculated using measured data.

[0055] Optionally, in step 5, the hyperparameters of the multi-task physical constraint deep learning model are automatically tuned using a Bayesian optimization algorithm based on Gaussian processes; a sliding time window online update mechanism is constructed to dynamically adjust the parameters of the multi-task physical constraint deep learning model to adapt to the watershed underlying surface and climate change.

[0056] Optionally, the method further includes, after step 5, performing uncertainty quantification and probability prediction, generating a prediction set using Monte Carlo Dropout technology, calculating the prediction mean and confidence interval of the data in the prediction set, and estimating the probability density distribution of power generation; and

[0057] Model interpretability analysis and decision support were conducted. The SHAP additive interpretation model was used to analyze the contribution of each input factor, such as effective rainfall, daily soil moisture, daily snowmelt water equivalent, daily evaporation, water level fluctuation, and inflow, to power generation at different spatiotemporal scales, and to generate spatiotemporal heat maps of key influencing factors.

[0058] A watershed hydropower generation prediction system based on multi-source data fusion and deep learning, the system comprising:

[0059] The data fusion module collects data from hydropower stations within the basin, and performs data fusion and spatiotemporal alignment to construct a coupled meteorological-hydrological-engineering dataset.

[0060] The effective rainfall calculation module constructs a grid-scale dynamic effective rainfall coefficient model and calculates the spatiotemporal sequence of effective rainfall considering spatiotemporal nonstationarity.

[0061] The feature engineering module performs multi-scale spatiotemporal feature engineering based on the meteorological-hydrological-engineering coupled dataset and the effective rainfall spatiotemporal sequence.

[0062] The deep learning prediction module establishes a multi-task physical constraint deep learning model with a shared feature extraction layer to predict hydropower generation.

[0063] The adaptive optimization module tunes and updates the hyperparameters of the multi-task physical constraint deep learning model online.

[0064] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to perform the steps of the method for predicting watershed hydropower generation based on multi-source data fusion and deep learning.

[0065] This invention provides a watershed hydropower generation prediction method based on multi-source data fusion and deep learning. It is applicable to the joint prediction of medium- and long-term and short-term power generation in complex watersheds containing multiple cascade hydropower stations. By constructing a multi-source heterogeneous data fusion framework encompassing meteorology, hydrology, and engineering, it proposes an effective rainfall calculation model considering the dynamic decay factor of reservoir storage status. A multi-scale spatiotemporal feature extraction module is designed, integrating CNN, Bi-LSTM, and self-attention mechanisms to capture spatial nonlinearity and long-term dependencies. A multi-task physical constraint deep learning framework is established to simultaneously predict power generation and available water volume, embedding the physical laws of hydropower conversion into the loss function. An adaptive online update strategy based on Bayesian optimization and Monte Carlo Dropout uncertainty quantification technology are employed to achieve probabilistic prediction and credibility assessment of power generation. SHAP interpretability analysis is integrated to achieve spatiotemporal analysis of factor contributions. This invention solves the problems of traditional methods failing to consider the spatial coupling effect of multiple reservoir groups within the watershed, the dynamic changes in engineering operation status, and the lack of physical mechanisms, significantly improving prediction accuracy and model robustness, and providing reliable decision support for hydropower scheduling in complex watersheds. Compared with existing technologies, this invention has the following technical advantages:

[0066] A dynamic effective rainfall calculation model was adopted: a reservoir energy storage state correction factor and a time decay factor were introduced to break through the limitations of the traditional static weighting method and accurately characterize the spatiotemporal dynamic utilization efficiency of rainfall by multiple reservoir groups.

[0067] A physical constraint deep learning framework was adopted: embedding physical consistency constraints based on the fundamental equations of hydropower generation reduced the model prediction error by 15%-20%;

[0068] It adopts a long-time input window design: using a 24-month history window, it exceeds the explicit lag order and fully learns long-range nonlinear dependency patterns. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0070] Figure 1 This is a flowchart of a method for predicting hydropower generation in a watershed based on multi-source data fusion and deep learning in a specific embodiment of the present invention;

[0071] Figure 2 This is a schematic diagram of the predicted hydropower generation during the test period in a specific embodiment of the present invention;

[0072] Figure 3 This is a schematic diagram of the structure of a watershed hydropower generation prediction system based on multi-source data fusion and deep learning in a specific embodiment of the present invention. Detailed Implementation

[0073] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0074] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0075] like Figure 1 As shown, Figure 1 This is a flowchart of a method for predicting hydropower generation in a river basin based on multi-source data fusion and deep learning, according to a specific embodiment of the present invention. The method includes:

[0076] Step 101, Multi-source heterogeneous data fusion and spatiotemporal alignment: Obtain meteorological observation data, hydrological monitoring data, engineering operation data and remote sensing inversion data of all hydropower stations in the basin, and construct a "meteorological-hydrological-engineering" coupled dataset with unified spatiotemporal resolution.

[0077] The multi-source heterogeneous data fusion and spatiotemporal alignment specifically include:

[0078] S1.1 Data Acquisition: Obtain the following data from N ≥ 10 hydropower stations within the basin:

[0079] (1) Meteorological data: hourly rainfall, temperature, humidity, wind speed, and radiation;

[0080] (2) Hydrological data: daily soil moisture, daily snowmelt water equivalent, daily evaporation;

[0081] (3) Engineering data: hourly reservoir water level, inflow, outflow, actual power generation, installed capacity, flood control limit water level, and design reservoir capacity;

[0082] (4) Remote sensing data: GPM satellite precipitation products, MODIS surface temperature and evapotranspiration products;

[0083] S1.2 Spatiotemporal Alignment: A triple nested grid system (coarse grid ≥ 0.1°, medium grid 0.05°, fine grid 0.01°) is used for spatial downscaling. Inverse distance weighted interpolation and terrain correction are used to achieve data assimilation from station to grid. In terms of time scale, wavelet packet decomposition and reconstruction are used to achieve unified temporal granularity alignment of data of different frequencies.

[0084] Step 102: Calculation of effective rainfall considering the dynamic energy storage state of the reservoir group: Based on the spatial topology relationship and installed capacity weight of the catchment area of ​​each hydropower station, a time decay factor and a reservoir energy storage state correction factor are introduced to construct a grid-scale dynamic effective rainfall coefficient model and calculate the spatiotemporal sequence of effective rainfall considering spatiotemporal nonstationarity.

[0085] The effective rainfall calculation model considering the dynamic energy storage state of the reservoir group is as follows:

[0086]

[0087] in:

[0088] For grid ( x , y At that moment t Effective rainfall;

[0089] Original rainfall data from the grid;

[0090] The spatial effective rainfall coefficient is calculated using the following formula:

[0091]

[0092]

[0093] In the formula, C i For the first i The installed capacity of each hydropower station; For the indicator function (whether the grid is in the first position) i Within the catchment area of ​​a power station, if so, the value is 1; otherwise, the value is 0. For the first i Power station t The availability of storage capacity at any given time. For the first i Each power station at any time t The actual storage capacity For the first i Total designed reservoir capacity of the power station;

[0094] β ( t The time decay factor is used, and the Gamma distribution function is employed to characterize the lag effect of rainfall-runoff. The calculation formula is as follows:

[0095]

[0096] In the formula, K The order of the Gamma distribution, for example, taking an integer value. K =3 means that the lag effect of 0 to 3 months is considered; ω k For the first k The weighting coefficients for the first lag can be determined using the maximum likelihood estimation method based on historical rainfall-runoff data. λ The shape parameters of the Gamma distribution are obtained by fitting historical runoff response curves using the least squares method; Γ( k () is the standard Gamma function;

[0097] This is a correction factor for the energy storage status, reflecting the amplification / inhibition effect of the overall regulation and storage capacity of the reservoir group on effective rainfall;

[0098]

[0099] In the formula, For the first i Each power station at any time t The reservoir water level (unit: m) is obtained in real time through the automatic hydrological monitoring and reporting system; For the first i Dead water level of a power station (unit: m); For the first i The normal water level of each power station (unit: m); min(·) is a function to take the minimum value, ensuring that the value is 1 when the water level exceeds the normal water level, thus avoiding numerical overflow.

[0100] Step 103, Multi-scale Spatiotemporal Feature Engineering: Perform multi-scale wavelet decomposition on effective rainfall, daily soil moisture, daily snowmelt water equivalent, daily evaporation, water level fluctuation (calculated from the daily water level monitoring data of the power station), and inflow, extract trend term, periodic term and residual term features, and construct spatiotemporal lag cross factor.

[0101] The multi-scale spatiotemporal feature engineering specifically includes:

[0102] S3.1 Perform three-level wavelet decomposition on the effective rainfall sequence to obtain the low-frequency trend component A3 and the high-frequency detail components D1, D2, and D3, which respectively represent the long-term trend, seasonal cycle, and random disturbance. After removing the random disturbance, the components are synthesized to obtain the effective rainfall sequence with random interference eliminated.

[0103] S3.2 Constructing the spatiotemporal lag cross-factor matrix F lag Considering the nonlinear lag relationship between rainfall and power generation:

[0104]

[0105]

[0106]

[0107]

[0108] In the formula, This represents the total effective rainfall in the basin. X × Y This represents the total number of grid cells in the watershed. This refers to the total inflow into the reservoir, which is the sum of the inflows from all hydropower stations. ( t ) is the first i Power station t Real-time inflow rate (unit: m³ / s); The average hydropower head for power generation in the basin is calculated by weighting the installed capacity. For the first i The installed capacity of each power station; For the first i Power station t Constantly generating water head; L R for The lag order is determined by the maximum mutual information coefficient (MIC), with a value ranging from 0 to 12 months. L Q The lag order of the flow is determined by the maximum mutual information coefficient (MIC), with a value ranging from 0 to 6 months. L HThe lag order of the water head is determined by the maximum mutual information coefficient (MIC), with a value ranging from 0 to 3 months.

[0109] Every moment t of For (1+ L R + L Q + L H A 3D vector, the feature matrix constructed over all time series. .

[0110] S3.3 Constructing Spatial Interaction Features: Calculating the spatial convolution feature between effective rainfall in the grid and available storage capacity of the upstream reservoir:

[0111]

[0112]

[0113]

[0114] In the formula, No. i The first power station t The spatial interaction feature vector at any given time (dimension 1×1) represents the combined impact of upstream water inflow and storage capacity on this power station. i Index for target power plant ( i =1,2,..., N ), j For grids or power plant indexes within the upstream influence domain; For the first i The upstream influence domain of a power station is defined as the directly connected upstream power station. j ; For spatial influence weighting, a combined calculation of flow direction and distance attenuation is used; For flow direction indication function, if the power station j Water flows towards the power station i but =1, otherwise =1 =0; For power station j to the power station i Length of water flow path (unit: km); D 0 represents the attenuation characteristic length, which is taken as 3 times the average reservoir spacing in the basin; For upstream power stations j At any moment t Available storage capacity; For power station j Total storage capacity; This refers to the availability of storage capacity.

[0115] S3.4 Constructing a unified spatiotemporal input tensor: This involves converting the watershed-level temporal features generated in S3.2 into a unified spatiotemporal input tensor. Power plant-level spatial interaction features generated with S3.3 By splicing the data, a three-dimensional input tensor is constructed. Among them: time dimension T The default value is 24 months to include at least two complete wet and dry water cycles and ensure sufficient preceding information; spatial dimension N The number of power stations in the basin; characteristic dimension D =2+ L R + L Q + L H .

[0116] Step 104: Construction of a multi-task physical constraint deep learning model: Establish a multi-task learning framework with a shared feature extraction layer. The main task is to predict hydropower generation, and the auxiliary task is to predict the available water volume in the basin. Embed a physical consistency constraint term based on the water balance equation and the power generation function into the loss function of the main task.

[0117] The structure of the physical constraint deep learning model is as follows:

[0118] The model is an end-to-end single-task deep learning architecture, and its input data is the three-dimensional input tensor constructed in step S3.4. Output the current time. t The predicted total power generation of the basin Specifically:

[0119] S4.1 Input Layer: Receives the three-dimensional input tensor X, and its construction follows the feature availability principle of the prediction stage, specifically including:

[0120] (1) Current moment t Total effective rainfall in the basin ;

[0121] (2) Known data at historical moments: t - 1, t - 2,…, t -L R Effective rainfall at any time t - 1,…, t -L Q Inbound traffic at any time t - 1,…, t -L H The constant head of the generator;

[0122] S4.2 Feature Extraction Layer: A 3-layer ConvLSTM is used to capture the space-time coupling mode, with each layer followed by a spatiotemporal attention mechanism. The attention weights are calculated as follows:

[0123]

[0124] Where σ is the sigmoid activation function. Here is the attention weight matrix. For bias vectors, This is the hidden state from the previous moment. For the current input features;

[0125] S4.3 Prediction Output Layer: A two-layer fully connected network is used, with the final hidden state of ConvLSTM. As input ( (Number of hidden nodes in ConvLSTM), output the current time step. t The predicted total power generation of the basin .

[0126] S4.4 Construct the main task loss function, which consists of two parts: deep learning error and physical consistency constraint:

[0127]

[0128]

[0129]

[0130] In the formula, L The main task loss function; The mean square error of power generation prediction; The physical consistency constraint term is constructed based on the fundamental equations of hydropower conversion. for t Actual observed values ​​of power generation at any given time. T This represents the total length of the time series. This represents the physical constraint weighting coefficient, with a value range of [0.1, 1.0]. η The comprehensive efficiency coefficient of the hydropower station is dimensionless, ranging from 0.75 to 0.92, and is calibrated based on historical data of the unit efficiency and head loss of each power station; g is the acceleration due to gravity, with a value of 9.81 m / s². 2 ; for t The average power generation flow of the basin at any given time is calculated using measured data. for t The average hydropower head of the basin at any given time is calculated using measured data.

[0131] Step 105, Adaptive Bayesian Optimization and Online Update: The model hyperparameters are automatically tuned using a Gaussian process-based Bayesian optimization algorithm; a sliding time window online update mechanism is constructed to dynamically adjust the model parameters to adapt to the watershed underlying surface and climate change.

[0132] The adaptive Bayesian optimization and online update are specifically as follows:

[0133] S5.1 Hyperparameter Optimization: Within the search space Θ={learning_rate, hidden_size, dropout_rate,λ}, a Gaussian process surrogate model is used to model the relationship between hyperparameters and validation loss functions. Five rounds of iterative optimization are performed using the expectation boosting (EI) strategy, with 20 configurations evaluated in each round.

[0134] S5.2 Online Update: Construct a sliding time window with a width of 24 months. When new data arrives, a transfer learning strategy is adopted to freeze the parameters of the underlying ConvLSTM and only update the parameters of the top fully connected layer and attention module. The update trigger condition is that the average relative error of prediction exceeds a threshold for three consecutive months. δ =10% or an extreme weather event may occur.

[0135] Step 106, Uncertainty Quantification and Probability Prediction: Generate a prediction ensemble using Monte Carlo Dropout technology, calculate the prediction mean and confidence interval, and estimate the probability density distribution of power generation.

[0136] Step 107, Model Interpretability Analysis and Decision Support: The SHAP additive interpretation model is used to analyze the contribution of each input factor to power generation at different spatiotemporal scales, and a spatiotemporal heat map of key influencing factors is generated to provide interpretable basis for scheduling decisions.

[0137] In a specific embodiment of the present invention, the method for predicting hydropower generation in a watershed based on multi-source data fusion and deep learning is used to predict hydropower generation, specifically including:

[0138] Step S1: Multi-source heterogeneous data fusion

[0139] (1) Collect data from 2015 to 2023 on 38 hydropower stations with an installed capacity of ≥50,000 kW in a certain province. The time resolution is hourly and daily, and the spatial resolution is 0.01° grid.

[0140] (2) Meteorological data came from the China Meteorological Administration's Land Surface Assimilation System (CLDAS), and remote sensing data came from GPM and MODIS products;

[0141] (3) Ensure spatial consistency of data through Kriging interpolation and terrain correction, and unify temporal granularity through wavelet packet decomposition.

[0142] Step S2: Calculation of Dynamic Effective Rainfall

[0143] (1) Use ArcGIS Pro 3.1 to extract the catchment areas of each hydropower station and calculate the spatial effective rainfall coefficient α;

[0144] (2) γ(t) was calculated based on the daily reservoir capacity availability rate, and the Gamma distribution parameters λ=0.85 and ω=[0.35,0.40,0.20,0.05] were calibrated using historical runoff data;

[0145] (3) Generate daily effective rainfall grid data from 2015 to 2023 with a spatial resolution of 0.05°.

[0146] Step S3: Multi-scale feature engineering and input construction

[0147] S3.1: Perform 3-level Daubechies wavelet decomposition on the effective rainfall and extract the A3 and D1-D3 components;

[0148] S3.2: Determining the lag order of rainfall through MIC analysis L R =6, flow lag order L Q =4, head lag order L H =2;

[0149] S3.3: Construct spatial interaction characteristics, the upstream influence domain Ω is determined according to the water flow path and propagation time;

[0150] S3.4: Constructing the 3D Input Tensor Among them: time dimension T The value is set to 24 months to include at least two complete wet and dry water cycles and ensure sufficient preceding information; spatial dimension N The number of power stations in the basin; characteristic dimension D =2+ L R + L Q + L H .

[0151] Step S4: Model Building and Training

[0152] (1) Input and output definition: The input is X (weather forecast for the past 24 months + time t), and the output is... t Total power generation in the basin at any time .

[0153] (2) Model structure: 3-layer ConvLSTM with 128 hidden layers, each layer followed by spatiotemporal attention; 2-layer fully connected output network; physical constraint loss weights. =0.4 (determined by Bayesian optimization).

[0154] (3) Training configuration: training set 2015-2020, validation set 2021, test set 2022-2023; AdamW optimizer, learning rate 0.0008, batch size=16, training for 100 epochs.

[0155] Step S5: Adaptive Optimization

[0156] (1) Bayesian optimization: The search space includes learning_rate∈[1e-4,5e-3], hidden_size∈{64,128,256}, dropout_rate∈[0.1,0.5], and λ∈[0.1,1.0]; the optimal combination of hyperparameters is determined after optimization.

[0157] (2) Online update: The sliding window width is 24 months. When MAPE>10% for 3 consecutive months or an extreme weather event occurs, transfer learning is triggered, the underlying parameters of ConvLSTM are frozen, and only the top-level network is fine-tuned.

[0158] Step S6: Result Verification and Analysis

[0159] like Figure 2 As shown, compared with the single-station prediction method, the present invention considers the time lag of upstream and downstream water volume and the coupling between power stations, and the Nash-Sutcliffe efficiency coefficient of the hydropower generation prediction result is improved from 0.78 to 0.89, which shows the superiority of the method of the present invention.

[0160] Accordingly, embodiments of the present invention also provide a watershed hydropower generation prediction system based on multi-source data fusion and deep learning, such as... Figure 3 The diagram shown is a structural schematic of the system. This watershed hydropower generation prediction system based on multi-source data fusion and deep learning includes the following modules:

[0161] Data fusion module 401 collects data from hydropower stations within the basin, performs data fusion and spatiotemporal alignment on the data, and constructs the meteorological-hydrological-engineering coupled dataset.

[0162] The effective rainfall calculation module 402 constructs a grid-scale dynamic effective rainfall coefficient model and calculates the spatiotemporal sequence of effective rainfall considering spatiotemporal nonstationarity.

[0163] The feature engineering module 403 performs multi-scale spatiotemporal feature engineering based on the meteorological-hydrological-engineering coupled dataset and the effective rainfall spatiotemporal sequence.

[0164] Deep learning prediction module 404 establishes a multi-task physical constraint deep learning model with a shared feature extraction layer to predict hydropower generation.

[0165] The adaptive optimization module 405 tunes and updates the hyperparameters of the multi-task physical constraint deep learning model online.

[0166] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0167] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it runs. Figure 1 The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.

[0168] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data provider to another website, computer, server, or data provider via wired or wireless means.

[0169] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and systems of the present invention, and are merely some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting hydropower generation in a river basin based on multi-source data fusion and deep learning, characterized in that, The method includes: Step 1: Collect data on hydropower stations within the basin, and perform data fusion and spatiotemporal alignment to construct a coupled meteorological-hydrological-engineering dataset; Step 2: Construct a grid-scale dynamic effective rainfall coefficient model and calculate the spatiotemporal sequence of effective rainfall considering spatiotemporal nonstationarity; Step 3: Perform multi-scale spatiotemporal feature engineering based on the meteorological-hydrological-engineering coupled dataset and the effective rainfall spatiotemporal sequence; Step 4: Establish a multi-task physical constraint deep learning model with a shared feature extraction layer to predict hydropower generation. Step 5: Optimize and update the hyperparameters of the multi-task physical constraint deep learning model online; In step 2, based on the spatial topological relationship and installed capacity weight of each hydropower station's catchment area, a time decay factor and a reservoir storage state correction factor are introduced to construct the grid-scale dynamic effective rainfall coefficient model, and the spatiotemporal sequence of effective rainfall considering spatiotemporal nonstationarity is calculated; wherein, the grid-scale dynamic effective rainfall calculation model considering the dynamic storage state of the reservoir group is as follows: in: For grid ( x , y At any moment t Effective rainfall; Original rainfall data from the grid; The spatial effective rainfall coefficient is calculated using the following formula: In the formula, C i For the first i The installed capacity of each hydropower station; As an indicator function, whether the grid is in the first position i Within the catchment area of ​​a power station, if it is, the value is 1; otherwise, the value is 0. For the first i Power station t The availability of storage capacity at any given time. For the first i Each power station at any time t The actual storage capacity For the first i Total designed reservoir capacity of the power station; β ( t The time decay factor is used, and the Gamma distribution function is employed to characterize the lag effect of rainfall-runoff. This is a correction factor for the energy storage status, reflecting the amplification / inhibition effect of the overall regulation and storage capacity of the reservoir group on effective rainfall; In step 3, multi-scale wavelet decomposition is performed on effective rainfall, daily soil moisture, daily snowmelt water equivalent, daily evaporation, water level fluctuation, and inflow. Trend, periodic, and residual features are extracted, and a spatiotemporal lag cross-factor is constructed, specifically including: Step 3.1: Perform three-level wavelet decomposition on the effective rainfall spatiotemporal sequence to obtain low-frequency trend component A3 and high-frequency detail components D1, D2, and D3, which respectively represent long-term trend, seasonal cycle, and random disturbance. After removing the random disturbance, the components are synthesized to obtain the effective rainfall sequence with random interference eliminated. Step S3.2: Construct the spatiotemporal lag crossover factor matrix F lag Considering the nonlinear lag relationship between rainfall and power generation: In the formula, This represents the total effective rainfall in the basin. X × Y This represents the total number of grid cells in the watershed. This refers to the total inflow into the reservoir, which is the sum of the inflows from all hydropower stations. ( t ) is the first i Power station t Real-time inflow rate (unit: m³ / s); The average hydropower head for power generation in the basin is calculated by weighting the installed capacity. For the first i The installed capacity of each power station; For the first i Power station t Constantly generating water head; L R for The lag order is determined by the maximum mutual information coefficient (MIC), with a value ranging from 0 to 12 months. L Q The lag order of the flow is determined by the maximum mutual information coefficient (MIC), with a value ranging from 0 to 6 months. L H The lag order of the water head is determined by the maximum mutual information coefficient MIC, with a value ranging from 0 to 3 months. Every moment t of For (1+ L R + L Q + L H A 3D vector, the feature matrix constructed over all time series. ; Step 3.3: Construct spatial interaction features and calculate the spatial convolution features of effective rainfall in the grid and available storage capacity of the upstream reservoir: In the formula, No. i The first power station t The spatial interaction feature vector at any given time represents the combined impact of upstream water inflow and storage capacity on this power station; i Index for the target power station, i =1,2,..., N ; j For grids or power plant indexes within the upstream influence domain; For the first i The upstream influence domain of a power station is defined as the directly connected upstream power station. j ; For spatial influence weighting, a combined calculation of flow direction and distance attenuation is used; For flow direction indication function, if the power station j Water flows towards the power station i but =1, otherwise =1 =0; For power station j to the power station i The length of the water flow path, in km; D 0 represents the attenuation characteristic length, which is taken as 3 times the average reservoir spacing in the basin; For upstream power stations j At any moment t Available storage capacity; For power station j Total storage capacity; This refers to the availability of storage capacity. Step 3.4: Construct a unified spatiotemporal input tensor, and convert the feature matrix constructed on the watershed-level time series. Interactive feature vectors of power plant-level space By splicing the data, a three-dimensional input tensor is constructed. Among them, the time dimension T The value is set to 24 months to include at least two complete wet and dry water cycles and ensure sufficient preceding information; spatial dimension N The number of power stations in the basin; characteristic dimension D =2+ L R + L Q + L H ; In step 4, a multi-task physical constraint deep learning model with a shared feature extraction layer is established. The main task predicts hydropower generation, and the auxiliary task predicts the available water volume in the basin. A physical consistency constraint term based on the water balance equation and the power generation function is embedded in the loss function of the main task. The multi-task physical constraint deep learning model is an end-to-end single-task deep learning architecture, and its input data is the three-dimensional input tensor. Output the current time. t The predicted total power generation of the basin .

2. The method for predicting hydropower generation in a watershed based on multi-source data fusion and deep learning according to claim 1, characterized in that, In step 1, the collected data from hydropower stations within the basin include meteorological observation data, hydrological monitoring data, engineering operation data, and remote sensing inversion data. The meteorological observation data includes hourly rainfall, temperature, humidity, wind speed, and radiation. The hydrological monitoring data includes daily soil moisture, daily snowmelt water equivalent, and daily evaporation. The engineering operation data includes hourly reservoir water level, inflow, outflow, actual power generation, installed capacity, flood control limit water level, and design reservoir capacity. The remote sensing inversion data includes GPM satellite precipitation products and MODIS surface temperature and evapotranspiration products. When fusing and aligning the data spatiotemporally, a triple-nested grid system is used for spatial downscaling. Inverse distance weighted interpolation and terrain correction are used to assimilate the data from stations to the grid. At the temporal scale, wavelet packet decomposition and reconstruction are used to achieve unified temporal granularity alignment of data from different frequencies, constructing a unified spatiotemporal resolution meteorological-hydrological-engineering coupled dataset.

3. The method for predicting hydropower generation in a watershed based on multi-source data fusion and deep learning according to claim 1, characterized in that, In step 4, the multi-task physical constraint deep learning model includes an input layer, a feature extraction layer, and a prediction output layer; The input layer receives the three-dimensional input tensor X, which is constructed following the feature availability principle of the prediction stage, specifically including: the current time step. t Total effective rainfall in the basin Known data for historical moments: t- 1, t- 2,…, tL R Effective rainfall at any time t- 1,…, tL Q Inbound traffic at any time t- 1,…, tL H The constant head of the generator; The feature extraction layer uses a 3-layer ConvLSTM to capture a space-time coupled pattern, with each layer followed by a spatiotemporal attention mechanism. The attention weights are calculated as follows: Where σ is the sigmoid activation function. Here is the attention weight matrix. For bias vectors, This is the hidden state from the previous moment. For the current input features; The prediction output layer uses a two-layer fully connected network, with the final hidden state of ConvLSTM. As input, where Output the number of hidden nodes in the ConvLSTM layer and the current time step. t The predicted total power generation of the basin ; The main task loss function is constructed, consisting of two parts: deep learning error and physical consistency constraint. In the formula, L The main task loss function; The mean square error of power generation prediction; The physical consistency constraint term is constructed based on the fundamental equations of hydropower conversion. for t Actual observed values ​​of power generation at any given time. T This represents the total length of the time series. These are the physical constraint weighting coefficients; η is the overall efficiency coefficient of the hydropower station, which is dimensionless; g is the acceleration due to gravity. for t The average power generation flow of the basin at any given time is calculated using measured data. for t The average hydropower head of the basin at any given time is calculated using measured data.

4. The method for predicting hydropower generation in a watershed based on multi-source data fusion and deep learning according to claim 1, characterized in that, In step 5, the hyperparameters of the multi-task physical constraint deep learning model are automatically tuned using a Bayesian optimization algorithm based on Gaussian processes; a sliding time window online update mechanism is constructed to dynamically adjust the parameters of the multi-task physical constraint deep learning model to adapt to the watershed underlying surface and climate change.

5. The method for predicting hydropower generation in a watershed based on multi-source data fusion and deep learning according to claim 1, characterized in that, The method also includes, after step 5, performing uncertainty quantification and probability prediction, generating a prediction set using Monte Carlo Dropout technology, calculating the prediction mean and confidence interval of the data in the prediction set, and estimating the probability density distribution of power generation; and Model interpretability analysis and decision support were conducted. The SHAP additive interpretation model was used to analyze the contribution of each input factor, such as effective rainfall, daily soil moisture, daily snowmelt water equivalent, daily evaporation, water level fluctuation, and inflow, to power generation at different spatiotemporal scales, and to generate spatiotemporal heat maps of key influencing factors.

6. A watershed hydropower generation prediction system based on multi-source data fusion and deep learning, characterized in that, The system includes: The data fusion module collects data from hydropower stations within the basin, and performs data fusion and spatiotemporal alignment to construct a coupled meteorological-hydrological-engineering dataset. The effective rainfall calculation module constructs a grid-scale dynamic effective rainfall coefficient model and calculates the spatiotemporal sequence of effective rainfall considering spatiotemporal nonstationarity. The feature engineering module performs multi-scale spatiotemporal feature engineering based on the meteorological-hydrological-engineering coupled dataset and the effective rainfall spatiotemporal sequence. The deep learning prediction module establishes a multi-task physical constraint deep learning model with a shared feature extraction layer to predict hydropower generation. The adaptive optimization module tunes and updates the hyperparameters of the multi-task physical constraint deep learning model online. The effective rainfall calculation module, based on the spatial topology and installed capacity weights of the catchment areas of each hydropower station, introduces a time decay factor and a reservoir storage state correction factor to construct the grid-scale dynamic effective rainfall coefficient model, and calculates the spatiotemporal sequence of effective rainfall considering spatiotemporal nonstationarity. The grid-scale dynamic effective rainfall calculation model considering the dynamic storage state of the reservoir group is as follows: in: For grid ( x , y At any moment t Effective rainfall; Original rainfall data from the grid; The spatial effective rainfall coefficient is calculated using the following formula: In the formula, C i For the first i The installed capacity of each hydropower station; As an indicator function, whether the grid is in the first position i Within the catchment area of ​​a power station, if it is, the value is 1; otherwise, the value is 0. For the first i Power station t The availability of storage capacity at any given time. For the first i Each power station at any time t The actual storage capacity For the first i Total designed reservoir capacity of the power station; β ( t The time decay factor is used, and the Gamma distribution function is employed to characterize the lag effect of rainfall-runoff. This is a correction factor for the energy storage status, reflecting the amplification / inhibition effect of the overall regulation and storage capacity of the reservoir group on effective rainfall; The feature engineering module performs multi-scale wavelet decomposition on effective rainfall, daily soil moisture, daily snowmelt water equivalent, daily evaporation, water level fluctuation, and inflow, extracting trend, periodic, and residual features, and constructing a spatiotemporal lag cross factor, specifically including: The effective rainfall spatiotemporal sequence is decomposed into three layers of wavelet decomposition to obtain low-frequency trend component A3 and high-frequency detail components D1, D2, and D3, which respectively represent long-term trend, seasonal cycle, and random disturbance. After removing the random disturbance, the components are synthesized to obtain the effective rainfall sequence with random interference eliminated. Construct the spatiotemporal lag cross factor matrix F lag Considering the nonlinear lag relationship between rainfall and power generation: In the formula, This represents the total effective rainfall in the basin. X × Y This represents the total number of grid cells in the watershed. This refers to the total inflow into the reservoir, which is the sum of the inflows from all hydropower stations. ( t ) is the first i Power station t Real-time inflow rate (unit: m³ / s); The average hydropower head for power generation in the basin is calculated by weighting the installed capacity. For the first i The installed capacity of each power station; For the first i Power station t Constantly generating water head; L R for The lag order is determined by the maximum mutual information coefficient (MIC), with a value ranging from 0 to 12 months. L Q The lag order of the flow is determined by the maximum mutual information coefficient (MIC), with a value ranging from 0 to 6 months. L H The lag order of the water head is determined by the maximum mutual information coefficient MIC, with a value ranging from 0 to 3 months. Every moment t of For (1+ L R + L Q + L H A 3D vector, the feature matrix constructed over all time series. ; Construct spatial interaction features and calculate the spatial convolutional features of effective rainfall in the grid and available storage capacity of the upstream reservoir: In the formula, No. i The first power station t The spatial interaction feature vector at any given time represents the combined impact of upstream water inflow and storage capacity on this power station; i Index for the target power station, i =1,2,..., N ; j For grids or power plant indexes within the upstream influence domain; For the first i The upstream influence domain of a power station is defined as the directly connected upstream power station. j ; For spatial influence weighting, a combined calculation of flow direction and distance attenuation is used; For flow direction indication function, if the power station j Water flows towards the power station i but =1, otherwise =1 =0; For power station j to the power station i The length of the water flow path, in km; D 0 represents the attenuation characteristic length, which is taken as 3 times the average reservoir spacing in the basin; For upstream power stations j At any moment t Available storage capacity; For power station j Total storage capacity; This refers to the availability of storage capacity. Construct a unified spatiotemporal input tensor, and convert the feature matrix constructed on the watershed-level time series. Interactive feature vectors of power plant-level space By splicing the data, a three-dimensional input tensor is constructed. Among them, the time dimension T The value is set to 24 months to include at least two complete wet and dry water cycles and ensure sufficient preceding information; spatial dimension N The number of power stations in the basin; characteristic dimension D =2+ L R + L Q + L H ; The deep learning prediction module establishes a multi-task physical constraint deep learning model with a shared feature extraction layer. The main task predicts hydropower generation, and the auxiliary task predicts the available water volume in the basin. A physical consistency constraint term based on the water balance equation and the power generation function is embedded in the loss function of the main task. The multi-task physical constraint deep learning model is an end-to-end single-task deep learning architecture, and its input data is the three-dimensional input tensor. Output the current time. t The predicted total power generation of the basin .

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it executes the steps of the method for predicting watershed hydropower generation based on multi-source data fusion and deep learning as described in any one of claims 1 to 5.

Images