A method, system and terminal for hydroelectric medium and long-term output prediction by fusing hydrological and meteorological big data

Abstract

The application discloses a kind of hydrology meteorological big data fusion hydroelectricity medium and long term output prediction method, system and terminal, it is related to electric power operation scheduling technical field, its technical scheme key points are: obtaining multi-source basic data and constructing watershed installed weighted area rainfall characteristics;A three-level coupled deep learning model composed of hydrology meteorological characterization layer, runoff response prediction layer and hydroelectricity output mapping layer is constructed, and the cascade reservoir constraint is embedded in the output mapping layer;According to the current hydrological state, generate state gating factor, dynamically adjust the weight and constraint strength inside the model;Nine serial physical consistency corrections are applied to the model output, and the probability output curve is obtained.The application realizes the uncertainty of meteorology along the physical chain, and the single model covers the whole flood and dry period through state gating and dynamic fusion of bucket experts, and the cascade constraint is embedded in the prediction process, and the whole process does not need manual intervention.

Description

Technical Field

[0001] This invention relates to the field of power operation and dispatching technology, and more specifically, to a method, system, and terminal for predicting the medium- and long-term output of hydropower by integrating hydrological and meteorological big data. Background Technology

[0002] Medium- and long-term forecasts of hydropower output are crucial for grid dispatching, power generation planning, and electricity market transactions. Current technologies primarily rely on historical runoff sequences, historical output sequences, or lumped hydrological models. These models estimate inflow based on rainfall, runoff, and confluence processes, then combine this with empirical dispatch rules to calculate output curves. Such methods are highly dependent on watershed underlying surface parameters, empirical manual calibration, and the accuracy of meteorological inputs, and are prone to systematic prediction biases under complex scenarios such as torrential rains, snowmelt runoff, and prolonged droughts. In recent years, deep learning models have been explored for hydropower output forecasting. Related research has made partial progress in areas such as multi-source precipitation data fusion, deep coupling of sub-seasonal precipitation and runoff, fusion of space-air-ground integrated satellite remote sensing and ground data, and deep learning-based estimation of hydrological baseflow and evaporation.

[0003] However, existing methods mostly focus on the in-situ optimization of single physical processes, making it difficult to unify multi-element forecasts from large meteorological models, runoff response, reservoir scheduling rules, cascade reservoir linkage, and hydropower output probability distribution into a single prediction framework. The uncertainty of meteorological forecasts cannot be physically mapped to the output probability distribution, and single models struggle to smoothly transition between flood season, dry season, and snowmelt season. Deep learning prediction results lack physical consistency guarantees, and output curves often exhibit time-series jumps, violations of water balance and unit ramping constraints, requiring extensive manual correction before they can be used for actual dispatch and electricity market transactions. Specific shortcomings of existing technologies include the following: First, the uncertainty of ensemble meteorological forecasts cannot be physically mapped to the output probability distribution across hydrological models and scheduling rules; existing solutions can only perform statistical post-processing at the output end, making the source of uncertainty untraceable. Second, single models cannot smoothly transition between flood season, dry season, and snowmelt season; prediction behavior under extreme hydrological conditions requires manual switching or independent modeling for different periods, increasing model maintenance costs and making jumps prone to occur at the switching points. Third, the cascade reservoir constraints operate as an external scheduling optimization module, separated from the prediction process. This makes it impossible to satisfy the cascade hydraulic coupling constraints during the output generation stage, requiring repeated verification and iteration after prediction, which is inefficient and easily disrupts prediction continuity. Fourth, deep learning prediction results lack physical consistency guarantees, and the output numerical curves have problems such as time series jumps, violations of water balance and unit ramping constraints, making them unsuitable for direct use in actual scheduling and power market transactions.

[0004] Therefore, researching and designing a method, system, and terminal for predicting medium- and long-term hydropower output that integrates hydrological and meteorological big data to overcome the above-mentioned shortcomings is an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method, system, and terminal for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data. This allows the ensemble forecast uncertainty of the meteorological big data model to pass through the three levels of hydrological state representation, runoff response prediction, and output mapping in a physically consistent manner. The final output is a multi-timescale probabilistic output curve that satisfies water balance, reservoir capacity boundary, cascade hydraulic connection, and unit ramping constraints, while also possessing physical interpretability. This enables unmanned intervention throughout the entire process from meteorological data to dispatch and trading decisions.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: Firstly, a method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data is provided, including the following steps: Acquire multi-source basic data, including meteorological big data model forecast data, river hydrological monitoring data, power station basic information and cascade correlation parameters; The multi-source basic data are fused to construct the watershed installed capacity weighted areal rainfall characteristics; A three-level coupled deep learning model is constructed, comprising a hydro-meteorological characterization layer, a runoff response prediction layer, and a hydropower output mapping layer connected in sequence, wherein the hydropower output mapping layer embeds cascade reservoir constraints. A state gating factor is generated based on the current hydrological state, and the internal weights and constraint strength of the three-level coupled deep learning model are dynamically adjusted through the state gating factor. Physical consistency correction is applied to the output of the three-level coupled deep learning model to obtain the final probability output curve.

[0007] Furthermore, the construction of the watershed installed weighted areal rainfall characteristics specifically involves: For each river basin, the weight of each city in the basin is calculated based on the proportion of installed capacity of each city within the basin. For each city, the weights of the city in all involved watersheds are summed and normalized to obtain the city's comprehensive weight. The rainfall data of each city are weighted and summed using the comprehensive weights to obtain the installed weighted areal rainfall characteristics of the watershed.

[0008] Furthermore, the runoff response prediction layer adopts a CNN-ResNet statistical downscaling coupled with a physically guided LSTM structure; The loss function of the physical-guided LSTM includes a prediction error term and a physical consistency penalty term. The physical consistency penalty term is the sum of squares of the difference between the inflow and outflow of the basin relative to the change in reservoir capacity during the prediction period.

[0009] Furthermore, the state gating factor is generated by combining flood season signals and the probability of extreme hydrological events; The comprehensive flood season signal is calculated by weighting the flow ratio and the degree of water level exceeding the warning level. The flow ratio is the ratio of the current flow to the average flow for the same period over many years. The degree of water level exceeding the warning level is the difference between the current water level and the flood limit water level divided by the difference between the normal storage water level and the flood limit water level. The probability of extreme hydrological events is output by an extreme hydrological event classifier based on satellite remote sensing and ensemble learning.

[0010] Furthermore, the three-level coupled deep learning model is equipped with a bucket expert prediction unit, which contains multiple LSTM expert models with the same structure but independent parameters. Each LSTM expert model is pre-trained for different prediction lead intervals. The bucket expert prediction unit generates fusion weights through a lightweight gated neural network. The input of the gated neural network is the current prediction lead and the state gate factor, and the output is the weight vector of each LSTM expert model. The weights are normalized by the softmax function to achieve smooth switching of a single model covering the entire prediction span.

[0011] Furthermore, the physical consistency correction is performed in a preset serial order, which is as follows: starting point smooth transition correction, output change rate limiting correction, cascade propagation time delay constraint correction, water balance constraint correction, reservoir capacity boundary constraint correction, long-term trend anchoring correction, expert interval boundary smoothing correction, sliding window smoothing correction, and dynamic range constraint correction. Among them, the starting point smooth transition correction is in the first position, used to eliminate the step between the initial moment of the model output and the measured value; the output change rate limiting correction is in the second position, used to limit the output change rate in adjacent time periods to not exceed the maximum ramp rate of the turbine unit.

[0012] Furthermore, the probability output curve includes three probability quantiles: P10, P50, and P90. These probability quantiles are directly generated by the quantile regression output layer in the three-level coupled deep learning model. The quantile regression output layer is trained using the pinball loss function, and there is no overlap between the probability quantiles.

[0013] Furthermore, the multi-source basic data also includes soil moisture data, which is derived from satellite remote sensing inversion or land surface model reanalysis products. After spatiotemporal interpolation and normalization, the soil moisture data is input as a supplementary feature into the hydro-meteorological characterization layer.

[0014] Secondly, a hydropower medium- and long-term output forecasting system integrating hydrological and meteorological big data is provided, including: The data acquisition module is used to acquire multi-source basic data, including meteorological big data model forecast data, river hydrological monitoring data, power station basic information and cascade correlation parameters; The feature construction module is used to fuse the multi-source basic data to construct the watershed installed weighted areal rainfall features; The model building module is used to build a three-level coupled deep learning model, which includes a hydro-meteorological characterization layer, a runoff response prediction layer, and a hydropower output mapping layer connected in sequence, wherein the hydropower output mapping layer embeds cascade reservoir constraints. The state gating module is used to generate a state gating factor based on the current hydrological state, and to dynamically adjust the internal weights and constraint strength of the three-level coupled deep learning model through the state gating factor; The correction output module is used to apply physical consistency correction to the output of the three-level coupled deep learning model to obtain the final probability output curve.

[0015] Thirdly, a computer terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data as described in any one of the first aspects.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention provides a method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data. It constructs a three-level coupled deep learning model comprising a hydrological and meteorological representation layer, a runoff response prediction layer, and a hydropower output mapping layer. Cascade reservoir constraints are embedded within the output mapping layer. Simultaneously, state gating factors are generated based on the current hydrological state to dynamically adjust the model's internal weights and constraint strengths. Finally, physical consistency correction is applied, achieving fully automated prediction from large-scale meteorological model forecasting to probabilistic output curves. This invention physically and consistently propagates the uncertainty of meteorological forecasts through the three levels of hydrological state representation, runoff response prediction, and output mapping, rather than performing statistical post-processing at the output. State gating factors enable the same model to automatically adjust information filtering preferences and constraint strengths under different hydrological conditions, eliminating the need for manual model switching. Physical consistency correction eliminates problems such as time-series jumps, water balance violations, and exceeding limits in the prediction curve. This invention solves the problems in the prior art where the uncertainty caused by the separation of weather forecasting, hydrological runoff, reservoir scheduling and output prediction cannot be physically transmitted, a single model cannot cover the entire span of flood and dry seasons, and the prediction results require a lot of manual correction before they can be used for scheduling and trading. It realizes unmanned intervention throughout the entire process from data input to decision-useful output.

[0017] 2. This invention calculates and normalizes rainfall characteristics based on the proportion of urban installed capacity, directly linking them to the actual hydropower generation capacity. This overcomes the limitation of traditional area-weighted areal rainfall, which fails to reflect differences in installed capacity, and improves the consistency between characteristic engineering and physical mechanisms. Furthermore, a physically guided LSTM is used in the runoff response prediction layer. Its loss function includes a physical consistency penalty term based on the water balance equation, forcing the model to learn the physical conservation relationship between watershed inflow, outflow, and reservoir capacity changes. Installed-capacity-weighted areal rainfall provides a more accurate rainfall-driven input, and the physically guided LSTM ensures that the runoff prediction process conforms to the water balance law, making the runoff prediction physically interpretable and traceable.

[0018] 3. This invention generates a state gating factor by integrating flood season signals and the probability of extreme hydrological events. The integrated flood season signal is calculated based on a weighted sum of the flow ratio and the degree of water level exceeding the warning level. The probability of extreme hydrological events is output through satellite remote sensing and an ensemble learning classifier, achieving accurate identification of different hydrological states such as flood season, dry season, and snowmelt season. Based on this, a bucket-based expert prediction unit is set up, containing multiple LSTM expert models pre-trained for different prediction lead intervals. A lightweight gated neural network is used to generate fusion weights with the current prediction lead and the state gating factor as input. This gating fusion mechanism elevates state identification from the input feature level to the model architecture level: the state gating factor dynamically adjusts the contribution weights of each expert model, making near-term predictions focus on real-time water conditions, while far-term predictions focus on seasonal trends, and achieving a smooth transition at the flood-dry season transition.

[0019] 4. This invention designs a sequential sequence for physical consistency corrections. These include: a smooth transition correction to eliminate initial step changes; a rate limiting correction to restrict the rate of ascent; a cascade propagation delay constraint correction to ensure upstream and downstream output timing conforms to hydraulic delays; water balance and reservoir capacity boundary constraint corrections to enforce physical conservation; long-term trend anchoring corrections to prevent deviations from climatic patterns at distant points; expert interval boundary smoothing corrections to eliminate bucket switching jumps; and sliding window smoothing and dynamic range constraint corrections to remove noise and limit physical boundaries. These nine corrections are not simply stacked independent post-processing steps, but rather sequentially coordinated according to causal logic: only by eliminating initial step changes can subsequent rate limiting have a correct benchmark; only by satisfying water balance can reservoir capacity boundary constraints have physical meaning. Based on this, the P10, P50, and P90 probability quantiles are directly generated through a quantile regression output layer, and a quantile loss function is used for training to ensure no quantile overlap, allowing the probability interval to directly originate from the physical propagation of front-end uncertainties. The final output of this invention achieves engineering-usable standards of temporal and physical consistency, constraint satisfaction, and reasonable probability intervals. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart from Embodiment 1 of the present invention; Figure 2 This is a 24-hour flow forecast curve from Embodiment 1 of the present invention. Figure 3 This is a 48-hour flow forecast curve from Embodiment 1 of the present invention. Figure 4 This is the flow forecast curve for the next 72 hours in Embodiment 1 of the present invention; Figure 5 This is a graph showing the variation of RMSE with prediction step size for different ablation models in Embodiment 1 of the present invention; Figure 6 This is a graph showing the variation of NSE with prediction step size for different ablation models in Embodiment 1 of the present invention; Figure 7 This is a graph showing the variation of physical residuals of different ablation models with prediction step size in Embodiment 1 of the present invention; Figure 8 This is a system block diagram in Embodiment 2 of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0022] Example 1: A method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data, such as... Figure 1 As shown, it includes the following steps: S1: Acquire multi-source basic data, including meteorological big data model forecast data, river hydrological monitoring data, power station basic information and cascade correlation parameters; S2: Integrate multi-source basic data to construct watershed installed weighted areal rainfall characteristics; S3: Construct a three-level coupled deep learning model, which includes a hydro-meteorological characterization layer, a runoff response prediction layer, and a hydropower output mapping layer connected in sequence, wherein the hydropower output mapping layer embeds cascade reservoir constraints. S4: Generate state gating factors based on the current hydrological state, and dynamically adjust the internal weights and constraint strength of the three-level coupled deep learning model through state gating factors; S5: Apply physical consistency correction to the output of the three-level coupled deep learning model to obtain the final probability output curve.

[0023] In step S1, multi-source basic data is first acquired, including meteorological big data model forecast data, river hydrological monitoring data, power station basic information, and cascade correlation parameters. The meteorological big data model forecast data originates from an AI-based meteorological big data model and a traditional ensemble numerical weather prediction system, containing hourly forecast fields for rainfall, temperature, humidity, wind speed, and other elements for the next 0 to 15 days, as well as corresponding ensemble member dispersion and probability forecast information. River hydrological monitoring data is collected in real-time through a network of hydrological stations deployed within the basin, including hourly water levels, flow rates, flow velocities, and historical flood records for each cross-section. Power station basic information includes parameters such as the installed capacity of each hydropower station, normal reservoir water level, flood control limit water level, dead water level, reservoir capacity curve, head-output relationship curve, maximum ramp rate of the generating units, and minimum technical output. Cascade correlation parameters include the topological connection relationship between upstream and downstream power stations, the river propagation time delay matrix, the inflow coefficient between sections, and the cascade joint scheduling rules.

[0024] In some preferred examples, the multi-source baseline data also includes soil moisture data. Soil moisture data is derived from satellite remote sensing inversion products, such as gridded soil moisture data from SMAP or ASCAT satellites, or from land surface model reanalysis products, such as hourly soil moisture data from ERA5-Land. The acquired raw soil moisture data is in a gridded format with a spatial resolution of approximately 0.1 to 0.25 degrees and a temporal resolution of hourly or daily. To match the watershed scale, the soil moisture data undergoes spatiotemporal interpolation: spatially, bilinear interpolation is used to interpolate the gridded data to the centroid of each watershed or the latitude and longitude of each city; temporally, cubic spline interpolation is used to align the data to the uniform time step required by the prediction model. After interpolation, the soil moisture data is normalized using the following formula: ; in, This represents the original soil moisture value. and These represent the minimum and maximum soil moisture values ​​for the site or grid point during historical periods. The normalized soil moisture data are used as supplementary features in the subsequent hydrometeorological characterization layer.

[0025] After acquiring the aforementioned multi-source basic data, quality control and time alignment are required for various data types. For meteorological large-scale model forecast data, outliers exceeding physical thresholds are removed, and the start time of all forecast members is unified to Coordinated Universal Time (UTC). For river hydrological monitoring data, jump data caused by equipment failures are removed, and missing measurements are interpolated using the correlation between adjacent stations. Power station basic information and cascade correlation parameters, as static data, are directly extracted from the power station operation management system and historical database, requiring no real-time updates but periodic verification. The time resolution of all data is unified to hourly; if the original data is daily or other resolutions, it is converted to an hourly sequence through linear interpolation or accumulation. After the above preprocessing, the multi-source basic data forms a structured dataset.

[0026] In step S2, after the multi-source basic data has undergone quality control and time alignment, this embodiment performs data fusion processing. The core objective is to construct a basin-wide installed capacity weighted areal rainfall characteristic. This characteristic differs from the traditional areal rainfall calculation method that weights areal rainfall by area or number of rain gauges. Instead, it uses the proportion of each city's installed capacity within the hydropower station's basin as the weight, thereby more accurately reflecting the contribution of rainfall to the actual power generation capacity of hydropower.

[0027] First, establish the mapping relationship of installed capacity weights between river basins and cities. Using river basin k as the key and the installed capacity percentage of each city i within that river basin as the value, cities across different river basins are weighted independently. That is, the formula for calculating the installed capacity percentage weight of city i in river basin k is: ; in, This represents the installed capacity of city i within the range of watershed k. Let be the set of cities contained within basin k, and let be the sum of the installed capacity of all cities within the basin. This weight reflects the relative importance of a single city's hydropower installed capacity within the overall power generation capacity of the basin.

[0028] Since a single city may span multiple watersheds, it is necessary to aggregate, accumulate, and normalize the data at the city level to obtain the comprehensive rainfall weight for each city. Let the set of watersheds to which city i belongs be denoted as . Then the unnormalized comprehensive weight of city i is The comprehensive rainfall weights for each city were obtained by normalizing the rainfall data. : ; Wherein, the denominator is the sum of the unnormalized comprehensive weights of all cities, and the numerator is the unnormalized comprehensive weight of city i. Iterate through all cities. By normalization, the sum of the comprehensive rainfall weights for all cities is 1.

[0029] Building upon this foundation, a standardized pipeline for city names is constructed. By establishing a fuzzy matching function based on string similarity, for a given original city name *s*, a weighted similarity function based on edit distance and a thesaurus is applied. In the standard city name set Matching the values ​​in the middle yields the standardized name: ; This step unifies and merges city names from different data sources, ensuring the accuracy of subsequent spatial associations.

[0030] Subsequently, a multi-level spatial association index was established. The mapping relationships between hydropower stations, rivers, and cities were extracted from the basic information table of hydropower stations, and the mapping relationships between monitoring stations and cities were extracted from river monitoring data, forming a three-level spatial association structure of monitoring stations, rivers, and cities. : ; in, This represents the m-th monitoring station. For the collection of monitoring stations; Represents the r-th river, For the collection of rivers; This represents the c-th city. A set of standard city names; This represents the spatial correlation distance weight from monitoring station m to city c. This structure allows hydrological observations from any monitoring station to be traced back to the meteorological data of the corresponding city, achieving spatial topological matching between meteorological grid points and watershed systems.

[0031] Construct a lookback window for a given reference time t. and forecast window Temporal characteristics of weighted areal rainfall at multiple scales within the basin. Watershed k at time... The weighted areal rainfall is defined as the installed weighted average of the rainfall forecast values ​​for each city within the basin: ; in, For the city i at any time The rainfall forecast value.

[0032] To overcome the systematic bias of single numerical forecasts in complex terrain regions, a multi-source precipitation fusion framework based on generative deep learning is adopted. This framework jointly encodes multi-source observed data and meteorological numerical forecast precipitation products, and the fusion formula is as follows: ; in, To merge in coordinates At any time The precipitation value; This is a multimodal fusion network based on an attention mechanism, with the following network parameters: ; For numerical weather prediction precipitation products, To estimate precipitation using radar, For satellite inversion of precipitation, This represents actual precipitation measured by a ground-based rain gauge. Network parameters. End-to-end training is performed by minimizing the joint loss function, which is: ; in, The mean squared error loss measures the difference between merged precipitation and observed precipitation. Deviation between; To counteract generation loss and improve the spatial texture realism of fused precipitation; As a gradient consistency loss, the spatial gradient of the fused precipitation is constrained to be consistent with the observed precipitation. , , These are the weighting coefficients for each loss term.

[0033] Simultaneously calculate the watershed weighted average humidity. : ; in, For the city i at any time The humidity forecast value and the short-term rainfall intensity index: ; in, This is the sliding step number, ranging from 0 to 11, corresponding to a sliding step of 15 minutes. The time within the sliding window; For the watershed k at time... The weighted areal rainfall. This indicator takes the maximum cumulative rainfall over the past 3 hours with a sliding step of 15 minutes, and is used to characterize the impact of short-term heavy rainfall.

[0034] Finally, power plant attribute features, provincial and regional meteorological features, dispatch margin features, historical trend features, time-coded features, and soil moisture supplementation features are extracted and concatenated into a complete feature vector. : ; in, The watershed installed weighted areal rainfall characteristic sequence, The watershed weighted average humidity series. This is a short-term rainfall intensity index sequence. These are the characteristics of the power station. The meteorological characteristics of the provincial region For scheduling margin characteristics, As a historical trend characteristic, For time-encoded features, It provides supplementary characteristics for soil moisture.

[0035] Soil moisture characteristics are based on daily 1-kilometer resolution soil moisture data generated by the fusion of multi-source remote sensing and ground observations, and are analyzed using a convolutional neural network spatial downscaling model. By integrating data from satellite microwave remote sensing of soil moisture, normalized difference vegetation index, land surface temperature, digital elevation model, and soil type, the areal mean soil moisture of each watershed is extracted: ; in, Let k be the spatial range of the watershed. Let k be the spatial area of ​​the watershed. The coordinate of time t Soil moisture was retrieved using satellite microwave remote sensing. Normalized Difference Vegetation Index; It refers to the surface temperature; For digital elevation models; This data represents soil type. The soil moisture characteristic serves as an underlying state feature characterizing the previous soil saturation level.

[0036] In step S3, based on the multi-source hydro-meteorological features constructed in step S2 and the hydrological state signal output in step S3, a three-level coupled deep learning model of hydrology, runoff and power output is constructed, including a hydro-meteorological characterization layer, a runoff response prediction layer and a hydropower output mapping layer connected in sequence, which is used to transform the meteorological big model forecast and multi-source hydro-meteorological data into medium and long-term hydropower output prediction results step by step.

[0037] The first level is the hydrometeorological characterization layer. This layer takes multi-source fusion precipitation ensemble forecasts, air temperature, humidity, evaporation, snowmelt water equivalent, soil moisture, watershed installed weighted areal rainfall, reservoir scheduling rules and characteristics, as well as flood season signals, dry season signals, and snowmelt runoff signals as inputs to extract future hydrological driving representations. These hydrological driving representations include precipitation driving representations, snowmelt recharge representations, evaporation deficit representations, previous soil saturation state representations, and scheduling constraint representations, used to characterize the impact of future meteorological disturbances on watershed inflow conditions.

[0038] The second level is the runoff response prediction layer. This layer takes as input the hydrological driving representation, historical flow, water level, anterior soil moisture, baseflow index, watershed storage index, and watershed static attribute embedding vector output from the hydrometeorological characterization layer, and predicts the runoff response under different lead times. The watershed static attribute embedding vector includes information on watershed area, average slope, river network density, soil saturated hydraulic conductivity, regulation capacity, and underlying surface type. In one implementation, the runoff response prediction layer adopts a coupled structure of CNN-ResNet statistical downscaling and physically guided LSTM. The CNN-ResNet statistical downscaling module downscales the multi-source fusion precipitation ensemble forecast from a coarse grid to a high-resolution grid, obtaining a fine precipitation field that matches the watershed scale. The physically guided LSTM takes the downscaled precipitation field and other hydrological driving representations as input to perform runoff time series prediction. The loss function of the physically guided LSTM includes a prediction error term and a physical consistency penalty term, where the physical consistency penalty term is the sum of squares of the deviations between the difference in watershed inflow and outflow relative to the change in reservoir capacity during the prediction period. Let the prediction period length be T, and the predicted inflow value at time t be... Outflow forecast value The change in storage capacity is Then the physical consistency penalty term The calculation formula is: ; in, , Let be the reservoir capacity at time t. This penalty term forces the model's output difference between inflow and outflow to numerically match the actual change in reservoir capacity, thereby compelling the model to learn the physical laws of water balance. The runoff response prediction layer is trained using a composite loss function consisting of runoff prediction error, Nash efficiency coefficient loss, flood peak error penalty term, and the aforementioned water balance constraint term, outputting runoff response prediction results at daily, weekly, and monthly scales.

[0039] The third level is the hydropower output mapping layer. This layer takes the runoff response prediction results, the current reservoir level or capacity, installed capacity, design head, available unit capacity, dispatch rules, upstream and downstream relationships of cascade reservoirs, and propagation time delays as inputs. It learns the nonlinear mapping relationship between inflow, reservoir capacity, head, unit capacity, and dispatch rules and hydropower output, and outputs the initial hydropower output prediction curve. For cascade reservoirs with upstream and downstream relationships, the outflow from the upstream reservoir, the power generation flow, or the discharge flow is transmitted to the downstream power station according to the propagation time delay and used as the constraint input for the downstream runoff response and output mapping. At the same time, the output of the cascade power station is jointly constrained according to the flood control limit level, normal storage level, dead water level, ecological discharge flow, upper and lower limits of power generation flow, and unit ramping capacity. The hydropower output mapping layer embeds cascade reservoir constraints, that is, the cascade hydraulic connection and reservoir capacity boundary are directly incorporated into the network structure within the model, rather than being externally checked after the prediction is completed. This embedded design allows the predicted curve to naturally satisfy the cascade hydraulic coupling constraint during the generation stage, eliminating the need for subsequent iterative adjustments.

[0040] To adapt to the error characteristics under different forecast lead times, a lead time expert prediction unit is set up in the three-level coupled model. Based on the forecast time lead, the forecast task is divided into near-end interval, mid-end interval, far-end interval, and ultra-far-end interval; each interval corresponds to a different prediction expert model. Unlike fixed-bucket hard switching, this invention generates fusion weights for each expert model based on the forecast lead and hydrological state signals. This allows near-end forecasts to focus on real-time hydrological conditions and short-term rainfall disturbances, while far-end forecasts focus on seasonal trends, baseflow status, and medium- to long-term meteorological background, thereby reducing the jumps in forecast curves at the transition between different lead time intervals. Hydrological state signals are not input as ordinary features but are used as state gating factors in the three-level coupled model. The essence of this design is to elevate state recognition from the input feature level to the model architecture level: the same three-level coupled model exhibits fundamentally different information filtering preferences, memory durations, and constraint strengths under different gating signals, enabling a single model architecture to cover extreme hydrological spans from floods to low water levels without the need to train and maintain independent models for different hydrological states. Meanwhile, the state gating factor directly participates in the dynamic generation of the expert fusion weights for the bucketing process, ensuring that the prediction curve at the transition between flood season and dry season is smooth and without jumps.

[0041] In the predictive inference phase, after determining the extrapolation starting point, the model sequentially generates hydrological and meteorological characterizations, runoff response prediction results, and initial hydropower output prediction curves based on the meteorological big data model forecasts, hydrological conditions, reservoir scheduling rules, and cascade reservoir constraints at each future prediction time. Subsequently, physical consistency corrections are applied to the initial prediction curves. These corrections include starting point smoothing, long-term trend anchoring, expert interval boundary smoothing, sliding window smoothing, dynamic range constraints, output change rate limiting, water balance constraints, reservoir capacity boundary constraints, and cascade propagation time lag constraints. For dry season predictions, a baseflow constraint module is introduced. This module uses accumulated rainfall, accumulated evapotranspiration, average soil moisture in the watershed, watershed static attributes, and historical low flow quantiles as inputs to predict the baseflow contribution during the dry season. This baseflow is then converted into a lower bound for sustainable output during the dry season, constraining the output of the output mapping layer and preventing abnormal fluctuations in the dry season prediction curve that do not conform to the baseflow replenishment pattern. After the above three-level coupled prediction and physical consistency correction, the output is a medium- and long-term hydropower output prediction curve that satisfies the constraints of hydrological process, reservoir scheduling rules and cascade joint operation, and serves as the basis for the subsequent generation of daily, weekly and monthly scale probability distributions.

[0042] In step S4, firstly, a flood season signal detection mechanism based on multi-indicator joint judgment is constructed. For key monitoring stations of major rivers, the ratio of the average flow over the most recent three days to the average flow over the previous seven days is calculated. Where m is the monitoring station index. The average flow rate over the past three days is denoted as... The calculation formula is as follows: ; in, Let be the flow rate at monitoring station m at time τ, and 3 × 96 represent the 288 15-minute time intervals over three days. The average flow rate for the previous seven days is denoted as... The calculation formula is as follows: ; Then the flow ratio .when A sudden increase in river flow is detected. Adaptive threshold. Based on the average of the historical traffic ratios for the same month at this station and standard deviation The calculation formula is dynamically set as follows: ; in, For adjustable sensitivity coefficient, This indicates the month in which time t falls.

[0043] Simultaneously check whether the water level at each monitoring station exceeds the warning level. Let the water level at monitoring station m at time t be... The warning water level is Then the water level warning indication function ,in, This is an indicator function that takes the value 1 if the condition is true and 0 otherwise.

[0044] Comprehensive flood season signals The result is obtained from the OR logic operation of sudden increase in flow rate and water level exceeding the warning level: ; in, This is a collection of all monitoring stations. This indicates a logical OR operation. When any monitoring station experiences a sudden increase in flow or a water level exceeding the warning level, the combined flood season signals are analyzed. It is true.

[0045] To further enhance the ability to identify extreme hydrological conditions, satellite remote sensing big data is introduced to assist in the identification of precursory conditions of extreme hydrological conditions. Multi-source remote sensing feature vectors are constructed. It includes the watershed normalized vegetation index. Surface temperature of the watershed and watershed snow water equivalent .in Acquired through microwave remote sensing inversion, it is used to identify precursors of snowmelt runoff. Simultaneously, a hydrological feature vector is constructed. This includes current flow rate, water level, and the aforementioned flow rate ratios. An ensemble learning model based on a weighted voting mechanism using random forest and gradient boosting tree is employed. right and Perform extreme hydrological state classification and identification, and output the probability of extreme hydrological events. : ; when When an extreme hydrological condition warning is triggered, among which, This is a preset probability threshold. This probability signal is used as an additional feature input to the subsequent prediction model.

[0046] Trend slope adaptive adjustment is achieved by calculating the time-series slope of the average daily power generation over the past seven days. Implementation. Let the average daily power generation before day d be... ,but: ; in, The slope is the least squares linear fit slope. According to... The output trend adjustment strategy for positive and negative values ​​and absolute values: If Then increase the recent weight, if If the rate of decrease is constrained, the normal mode is maintained.

[0047] The state gating factor is derived from the comprehensive flood season signal. and the probability of extreme hydrological events Co-generation. Specifically, the comprehensive flood season signal is used as a hard-decision gating factor to directly trigger the model to enter flood season mode; the probability of extreme hydrological events is used as a soft-probability feature, and is input into the state gating module along with the comprehensive flood season signal to finely adjust the weight allocation within the model. State gating factor It is a multidimensional vector whose dimensions correspond to the number of adjustable parameters in the three-level coupled model, including four dimensions: internal weight, hysteresis weight, constraint strength, and probability interval width.

[0048] A bucketed expert prediction unit is set up inside the three-level coupled deep learning model. The bucketed expert prediction unit contains multiple LSTM expert models with identical structures but independent parameters, each pre-trained for a different prediction lead interval. Based on the prediction time lead, the prediction task is divided into near-end interval, mid-end interval, far-end interval, and ultra-far-end interval, with each interval corresponding to an expert model. The bucketed expert prediction unit generates fusion weights through a lightweight gated neural network. The input of the gated neural network is the current prediction lead Δt and the state gating factor. The output is the weight vector of each LSTM expert model. Where N is the number of expert models. The weight vector is normalized using the softmax function: ; in, Let be the j-th raw score output by the gated neural network. The normalized weights satisfy... The final prediction result is a weighted sum of the outputs of each expert model: ; in, This represents the prediction output of the j-th expert model at time t. This dynamic fusion mechanism allows near-term predictions to focus on real-time hydrological conditions and short-term rainfall disturbances, while long-term predictions focus on seasonal trends, baseflow conditions, and medium- to long-term meteorological background, thereby reducing the jumps in prediction curves at the transition between different lead time intervals.

[0049] Hydrological state signals are not input as ordinary features alone, but rather act as state gating factors on the three-level coupled model. The essence of this design is to elevate state recognition from the input feature level to the model architecture level: the same three-level coupled model exhibits fundamentally different information filtering preferences, memory durations, and constraint strengths under different gating signals, enabling a single model architecture to cover extreme hydrological spans from flood to low water levels without the need to train and maintain independent models for different hydrological states. Simultaneously, the state gating factor directly participates in the dynamic generation of bucket expert fusion weights, ensuring a smooth and abrupt transition in the prediction curve at the flood-low water transition point. This effect is impossible to achieve when state recognition and bucket modeling are used independently; it can only be produced when the two work together using a gating fusion mechanism.

[0050] Specifically, when the hydrological conditions are identified as flood season, sudden flow surge, or water level exceeding warning levels, the state gating factor increases the weights corresponding to future rainfall forecasts, short-term cumulative rainfall, real-time flow growth rate, reservoir capacity adequacy, and flood discharge scheduling rules. When the conditions are identified as dry season or high evaporation and low rainfall, the weights corresponding to evaporation deficit, consecutive days of low rainfall, low flow percentile, baseflow index, and ecological discharge constraints are increased. When the conditions are identified as snowmelt recharge, the weights corresponding to snow water equivalent, temperature crossing zero change, and upstream inflow growth rate are increased. Through this dynamic adjustment, the three-level coupled model can adaptively adjust its internal behavior, thereby maintaining optimal predictive performance under different hydrological conditions.

[0051] In step S5, after the three-level coupled deep learning model generates the initial hydropower output prediction curve, a physical consistency correction is applied to the initial prediction curve to eliminate problems such as time series jumps, violations of water balance, exceeding the unit's ramp-up capability, and deviations from long-term patterns, so that the output curve meets the engineering usability standard that can be directly used for dispatching and electricity market trading. Subsequently, the corrected output sequence is aggregated across multiple time scales, and a probabilistic output curve is generated through a quantile regression output layer.

[0052] The physical consistency correction is performed in a preset sequential order and includes nine corrections: starting point smooth transition correction, output change rate limiting correction, cascade propagation time delay constraint correction, water balance constraint correction, reservoir capacity boundary constraint correction, long-term trend anchoring correction, expert interval boundary smoothing correction, sliding window smoothing correction, and dynamic range constraint correction. Each correction is performed based on the previous one, forming a serial and coordinated correction pipeline. The absence of any one of these corrections will render the output unusable.

[0053] The initial smooth transition correction is placed first to eliminate the step difference between the model output's initial time and the measured value. Let the initial time of the model output be... The predicted output value at that moment is The measured output value is If there is a discrepancy between the two... Then, exponential decay compensation is applied to several time steps in the initial segment of the predicted curve, allowing the predicted curve to smoothly transition from the measured value to the model-predicted trajectory. Specifically, the corrected output value... In the initial transition window The internal calculation is as follows: ; in, The initial predicted output value of the three-level coupled deep learning model; For the transition window length, The decay time constant is set according to the actual response characteristics of the power plant.

[0054] The output change rate limiting correction is the second parameter, used to limit the output change rate between adjacent time periods from exceeding the turbine unit's maximum ramp rate. Let the turbine unit's maximum ramp rate be... The unit is megawatts per minute. The maximum permissible change in output for two consecutive 15-minute intervals is... During calibration, the process begins by checking each point on the predicted curve from the second point onwards. If the output difference between the current point and the previous moment exceeds [a certain threshold], [the calibration is performed]. Then the output at the current point will be limited to the output at the previous moment plus or minus. ,Right now: ; in, This is the output value after rate limiting correction.

[0055] Cascade propagation time delay constraint correction is used to ensure that the output timing relationship between upstream and downstream power plants conforms to the physical delay of water flow propagation. Let the outflow rate of the upstream power plant be u. The inflow of the downstream power station d section is River propagation time lag is Then the inflow of the downstream power station It should meet the following requirements: ; During calibration, the upstream and downstream output timing is adjusted according to the cascade propagation time delay matrix to align the downstream output sequence with the upstream outflow in time.

[0056] Water balance constraint correction is used to force the prediction curve to satisfy the reservoir water balance equation. Assume the reservoir is in the time period... The internal storage capacity changes are Inbound flow is Outbound flow is The water balance equation is then: ; During calibration, the predicted output curve is converted back into the outflow rate and substituted into the water balance equation for verification. If the deviation exceeds the allowable range, the output value is adjusted iteratively to make the equation approximately valid.

[0057] Reservoir capacity boundary constraint correction is used to ensure that the reservoir capacity remains at the dead water level throughout the forecasting process. and normal water storage capacity Between, and during the flood season, the reservoir capacity corresponding to the flood limit level. If the predicted reservoir capacity exceeds the limit, the capacity will be brought back to the allowable range by adjusting the power output or the discharge flow.

[0058] Long-term trend anchoring correction is used to prevent long-term forecasts from deviating from climatological patterns. Let the historical multi-year average power output sequence for the same period be... The forecast for the far end should fluctuate around the climatological mean. During correction, a slow regression term is applied to the output values ​​in the latter half of the forecast window to gradually bring them closer to the climatological mean. The regression strength increases with the increase of the forecast lead time.

[0059] Expert interval boundary smoothing correction is used to eliminate artificial jumps at the switching points of bucketed expert models. Near the boundary point between adjacent expert models, the predicted values ​​on both sides are weighted and averaged. The weighting coefficient changes linearly with the distance from the boundary point, thus achieving a smooth transition.

[0060] Sliding window smoothing correction uses a moving average filter to remove high-frequency noise from the predicted curve. Let the sliding window length be L, then the corrected output value is the arithmetic mean of the original predicted values ​​within the window: ; Dynamic range constraint correction is used to limit the final output value within a reasonable physical range, i.e., not lower than the unit's minimum technical output. And not exceeding the installed capacity Values ​​outside the range are truncated to the boundary values.

[0061] After the above nine serial corrections, the prediction curve reaches the engineering standard of being time- and physical consistent, constraint satisfied, and ready for direct deployment.

[0062] Next, the corrected power output prediction sequence is aggregated by day, week, and month to obtain power output samples at different time scales. The probabilistic power output curve includes three probability quantiles: P10, P50, and P90. These probability quantiles are directly generated by the quantile regression output layer in a three-level coupled deep learning model, rather than being obtained through statistical post-processing after prediction. The quantile regression output layer is trained using a quantile loss function, also known as the pinball loss function. For a given quantile level... The quantile loss function is defined as: ; in, This represents the actual effort output. This represents the output value predicted by the model at the q-th quantile. When q = 0.1, it corresponds to the P10 quantile; when q = 0.5, it corresponds to the P50 quantile; and when q = 0.9, it corresponds to the P90 quantile. During training, the sum of the losses for the three quantiles is optimized simultaneously, and a quantile crossover penalty term is applied to ensure that there is no crossover between quantiles, i.e., the condition is always met. .

[0063] Furthermore, historical forecast errors are calibrated by grouping them according to the flood season, dry season, snowmelt season, and high evaporation and low rainfall period. When precipitation forecast dispersion is large, water levels exceed warning levels, or there is a sudden increase in flow, the probability interval is widened; when baseflow is stable during the dry season and reservoir capacity is clear, the probability interval is narrowed. This dynamic interval adjustment ensures that the uncertainty range of the probabilistic output curve matches the actual risk, thus providing a reliable decision-making basis for grid dispatching and electricity market transactions.

[0064] Validation example: Ablation validation of runoff response prediction layer in the Dadu River Basin.

[0065] We selected hourly ERA5 weather data from Sichuan Province and hydrological monitoring data from the Dadu River Basin, with Luding (II) station as the target station and Danba station as the upstream proxy station. The last 14 days were used as the test set. We compared and verified the data from five dimensions: data settings, ablation model, prediction results, gain analysis, and typical curves.

[0066] 1. Experimental data and ablation configuration.

[0067] The experimental data sources and preprocessing settings are shown in Table 1.

[0068] Table 1 Data Sources and Preprocessing Settings

[0069] To quantify the independent contributions of weather data, multi-scale weather coding, and physical consistency constraints to the runoff response prediction layer, five types of control configurations were constructed: persistent baseline, historical flow / water level model, original weather feature model, multi-scale weather coding model, and simplified physical constraint model, as shown in Table 2.

[0070] Table 2 Ablation Experiment Configuration

[0071] 2. Ablation experiment.

[0072] Evaluation metrics include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), Nash efficiency coefficient (NSE), and physical residual RMSE, with RMSE and physical residual RMSE being the main comparison metrics. RMSE results for different prediction step sizes are shown in Table 3.

[0073] Table 3 Comparison of RMSE for different ablation models

[0074] Table 3 shows that M2, which incorporates original weather features, reduces medium- and long-term prediction errors compared to M1, which only uses historical flow / water level data. Furthermore, by adding multi-scale weather coding, M3 improves the RMSE of M2 by 28.59%, 40.40%, and 45.25% at prediction steps of 24h, 48h, and 72h, respectively. This indicates that cumulative precipitation, short- and long-window residuals, and multi-scale statistics on temperature, cloud cover, and radiation can enhance the model's ability to represent future runoff changes, thus supporting the effectiveness of the CNN-ResNet statistical downscaling structure in the runoff response prediction layer.

[0075] 3. Ablation gain and curve verification.

[0076] To further illustrate the independent contributions of each component, statistics were compiled on RMSE improvement and simplified physical residual improvement, and the results are shown in Table 4.

[0077] Table 4 Ablation Gain Analysis

[0078] Table 4 shows that the original weather characteristics improved the RMSE of M1 by 17.01% to 22.14%, the multi-scale weather coding improved the RMSE of M2 by 28.59% to 45.25%, and the physical residual RMSE improvement of M3 relative to the persistent baseline was 3.93% to 5.59%. Since the experimental period mainly occurred during the winter dry season, the flow changes were relatively stable, and the persistent baseline had a strong inertial advantage in RMSE. However, the persistent baseline cannot characterize the cumulative impact of meteorological disturbances on future runoff, while M3 can reflect the gains of multi-scale weather coding in both error and physical residual dimensions.

[0079] Furthermore, this embodiment uses flow forecast curves for the next 24h, 48h, and 72h, as well as curves showing the changes in RMSE, NSE, and physical residuals with the forecast step size, to demonstrate typical results, as shown below. Figures 2 to 7As shown in the figure, black, blue, green, orange, purple, and red represent the actual flow, M0 persistent baseline, M1 historical flow / water level model, M2 original weather feature model, M3 multi-scale weather coding model, and M4 simplified physical constraint model, respectively. The curve comparison shows that multi-scale weather coding can significantly reduce the bias expansion of M2 in medium- and long-term forecasts and keep the physical residual relatively stable as the forecast step size increases. The simplified physical constraint model M4 did not consistently outperform M3 under the current data conditions, mainly because the experimental data lacked real reservoir capacity curves, inflow / outflow flow, and power plant scheduling information. In the actual system of this invention, the physically guided LSTM uses complete water balance constraints and cascade reservoir constraints, which can avoid the shift of the surrogate residual to the training target.

[0080] Example 2: A hydropower medium- and long-term output prediction system integrating hydrological and meteorological big data. This system is used to implement the hydropower medium- and long-term output prediction method integrating hydrological and meteorological big data described in Example 1, such as... Figure 8 As shown, it includes a data acquisition module, a feature construction module, a model construction module, a state gating module, and a correction output module.

[0081] The system comprises the following modules: a data acquisition module for acquiring multi-source basic data, including meteorological big data model forecasts, river hydrological monitoring data, power station basic information, and cascade correlation parameters; a feature construction module for fusing the multi-source basic data to construct the watershed installed capacity weighted areal rainfall characteristics; a model construction module for constructing a three-level coupled deep learning model, comprising a hydro-meteorological characterization layer, a runoff response prediction layer, and a hydropower output mapping layer connected in sequence, with the hydropower output mapping layer embedding cascade reservoir constraints; a state gating module for generating state gating factors based on the current hydrological state and dynamically adjusting the internal weights and constraint strength of the three-level coupled deep learning model through these state gating factors; and a correction output module for applying physical consistency correction to the output of the three-level coupled deep learning model to obtain the final probabilistic output curve.

[0082] Example 3: A computer terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data as described in Example 1.

[0083] Working Principle: This invention constructs a three-level coupled deep learning model comprising a hydro-meteorological characterization layer, a runoff response prediction layer, and a hydropower output mapping layer. Cascade reservoir constraints are embedded within the output mapping layer. Simultaneously, state gating factors are generated based on the current hydrological state to dynamically adjust the model's internal weights and constraint strength. Finally, physical consistency correction is applied, achieving fully automated prediction across the entire chain from large-scale meteorological model forecasting to probabilistic output curves. This invention physically and consistently propagates the uncertainty of meteorological forecasts through the three levels of hydrological state characterization, runoff response prediction, and output mapping, rather than performing statistical post-processing at the output. State gating factors enable the same model to automatically adjust information filtering preferences and constraint strength under different hydrological conditions, eliminating the need for manual model switching. Physical consistency correction eliminates problems such as time-series jumps, water balance violations, and exceeding limits in the prediction curve. This invention solves the problems in the prior art where the uncertainty caused by the separation of weather forecasting, hydrological runoff, reservoir scheduling and output prediction cannot be physically transmitted, a single model cannot cover the entire span of flood and dry seasons, and the prediction results require a lot of manual correction before they can be used for scheduling and trading. It realizes unmanned intervention throughout the entire process from data input to decision-useful output.

[0084] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0085] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0086] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0088] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. 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 medium- and long-term hydropower output by integrating hydrological and meteorological big data, characterized in that, Includes the following steps: Acquire multi-source basic data, including meteorological big data model forecast data, river hydrological monitoring data, power station basic information and cascade correlation parameters; The multi-source basic data are fused to construct the watershed installed capacity weighted areal rainfall characteristics; A three-level coupled deep learning model is constructed, comprising a hydro-meteorological characterization layer, a runoff response prediction layer, and a hydropower output mapping layer connected in sequence, wherein the hydropower output mapping layer embeds cascade reservoir constraints. A state gating factor is generated based on the current hydrological state, and the internal weights and constraint strength of the three-level coupled deep learning model are dynamically adjusted through the state gating factor. Physical consistency correction is applied to the output of the three-level coupled deep learning model to obtain the final probability output curve.

2. The method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data according to claim 1, characterized in that, The construction of the watershed installed weighted areal rainfall characteristics is specifically as follows: For each river basin, the weight of each city in the basin is calculated based on the proportion of installed capacity of each city within the basin. For each city, the weights of the city in all involved watersheds are summed and normalized to obtain the city's comprehensive weight. The rainfall data of each city are weighted and summed using the comprehensive weights to obtain the installed weighted areal rainfall characteristics of the watershed.

3. The method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data according to claim 1, characterized in that, The runoff response prediction layer adopts a coupled structure of CNN-ResNet statistical downscaling and physically guided LSTM. The loss function of the physical-guided LSTM includes a prediction error term and a physical consistency penalty term. The physical consistency penalty term is the sum of squares of the difference between the inflow and outflow of the basin relative to the change in reservoir capacity during the prediction period.

4. The method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data according to claim 1, characterized in that, The state gating factor is generated by combining flood season signals and the probability of extreme hydrological events. The comprehensive flood season signal is calculated by weighting the flow ratio and the degree of water level exceeding the warning level. The flow ratio is the ratio of the current flow to the average flow for the same period over many years. The degree of water level exceeding the warning level is the difference between the current water level and the flood limit water level divided by the difference between the normal storage water level and the flood limit water level. The probability of extreme hydrological events is output by an extreme hydrological event classifier based on satellite remote sensing and ensemble learning.

5. The method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data according to claim 1, characterized in that, The three-level coupled deep learning model has a bucket expert prediction unit inside. The bucket expert prediction unit contains multiple LSTM expert models with the same structure but independent parameters. Each LSTM expert model is pre-trained for different prediction lead intervals. The bucket expert prediction unit generates fusion weights through a lightweight gated neural network. The input of the gated neural network is the current prediction lead and the state gate factor, and the output is the weight vector of each LSTM expert model. The weights are normalized by the softmax function to achieve smooth switching of a single model covering the entire prediction span.

6. The method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data according to claim 1, characterized in that, The physical consistency correction is performed in a preset serial order, which is as follows: starting point smooth transition correction, output change rate limiting correction, cascade propagation time delay constraint correction, water balance constraint correction, reservoir capacity boundary constraint correction, long-term trend anchoring correction, expert interval boundary smoothing correction, sliding window smoothing correction, and dynamic range constraint correction. Among them, the starting point smooth transition correction is in the first position, used to eliminate the step between the initial moment of the model output and the measured value; the output change rate limiting correction is in the second position, used to limit the output change rate in adjacent time periods to not exceed the maximum ramp rate of the turbine unit.

7. The method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data according to claim 1, characterized in that, The probability output curve includes three probability quantiles: P10, P50, and P90. These probability quantiles are directly generated by the quantile regression output layer in the three-level coupled deep learning model. The quantile regression output layer is trained using the pinball loss function, and there is no overlap between the probability quantiles.

8. The method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data according to claim 1, characterized in that, The multi-source basic data also includes soil moisture data, which is derived from satellite remote sensing inversion or land surface model reanalysis products. After spatiotemporal interpolation and normalization, the soil moisture data is input as a supplementary feature into the hydro-meteorological characterization layer.

9. A hydropower medium- and long-term output prediction system integrating hydrological and meteorological big data, characterized in that, include: The data acquisition module is used to acquire multi-source basic data, including meteorological big data model forecast data, river hydrological monitoring data, power station basic information and cascade correlation parameters; The feature construction module is used to fuse the multi-source basic data to construct the watershed installed weighted areal rainfall features; The model building module is used to build a three-level coupled deep learning model, which includes a hydro-meteorological characterization layer, a runoff response prediction layer, and a hydropower output mapping layer connected in sequence, wherein the hydropower output mapping layer embeds cascade reservoir constraints. The state gating module is used to generate a state gating factor based on the current hydrological state, and to dynamically adjust the internal weights and constraint strength of the three-level coupled deep learning model through the state gating factor; The correction output module is used to apply physical consistency correction to the output of the three-level coupled deep learning model to obtain the final probability output curve.

10. A computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a method for predicting medium- and long-term hydropower output by integrating hydrological and meteorological big data as described in any one of claims 1-8.

Images