Reservoir scheduling rolling simulation method and device coupling data driving and physical constraint

Abstract

This invention discloses a rolling simulation method and apparatus for reservoir scheduling that couples data-driven and physical constraints. The method includes: acquiring historical reservoir operation data, including reservoir water level, inflow, outflow, and time information; using the reservoir capacity at the beginning of a time period, inflow, and number of days in the year as feature variables, and the reservoir capacity at the end of the time period as the target variable, training a scheduling simulation model using a machine learning algorithm; during the model prediction process, introducing water balance equations and multiple physical constraints on reservoir operation, and using a rolling prediction method, taking the prediction output of the previous time period as the input of the next time period, simulating and generating a long series of outflow processes period by period. This invention overcomes the shortcomings of complex traditional physical models and poor adaptability of empirical models, achieving high-precision and high-efficiency simulation of large reservoir scheduling processes. It can be used for hydrological situation impact analysis, scheduling scheme evaluation, and extreme event early warning, providing core technical support for intelligent reservoir scheduling decision-making.

Description

Technical Field

[0001] This invention relates to the field of reservoir scheduling technology in water conservancy engineering, specifically to a rolling simulation method and apparatus for reservoir scheduling that couples data-driven and physical constraints. Background Technology

[0002] Reservoir scheduling is a core aspect of water conservancy project operation and management. Its core task is to maximize the comprehensive benefits of multiple objectives, including flood control, power generation, water supply, navigation, and ecological protection, through the scientific regulation of reservoir outflow. Traditional scheduling methods mainly rely on hydrological and hydraulic models based on physical mechanisms or empirical scheduling chart methods based on regulations. Physical models have clear mechanisms and interpretability, but they are usually complex to construct, computationally inefficient, and heavily dependent on a large number of precisely calibrated parameters, making them difficult to promote and apply in practical operations. While empirical scheduling chart methods are simple in structure and easy to operate, they lack flexibility and are difficult to adapt to the increasingly complex and volatile hydrological situations and the growing diversified scheduling needs (such as flood control, power generation, navigation, and ecological protection) under the background of climate change, lacking flexibility and adaptability.

[0003] In recent years, with the rapid development of artificial intelligence technology, machine learning methods have demonstrated powerful nonlinear fitting and high-dimensional time-series data processing capabilities, and have gradually been introduced into the field of reservoir scheduling. These data-driven methods can automatically extract scheduling patterns from historical operational data, circumventing the stringent requirements of traditional physical models on mechanistic structures. However, most existing machine learning applications still have significant limitations: they directly establish end-to-end prediction models from input to outflow, neglecting the most fundamental water balance principle and other key physical constraints in reservoir systems (such as water level fluctuation limits, extreme values ​​of outflow, and power station output range). The result is often that the model performs well in statistical indicators, but the generated scheduling process cannot be implemented in the actual physical system, or clearly violates operational procedures, leading to difficulties in technology implementation and low engineering practicality.

[0004] Therefore, in the context of promoting smart water conservancy and digital twin watershed construction, there is an urgent need to develop a new reservoir scheduling simulation technology that can fully leverage the powerful data mining and prediction capabilities of machine learning, while also embedding the physical mechanisms and operational rules of the water conservancy field into the modeling process in a computable and constrained manner. Research in this direction will not only help improve the intelligence and adaptability of scheduling decisions, but also have significant practical implications for ensuring watershed water security and achieving efficient water resource utilization. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a GBM-based rolling simulation method and device for reservoir scheduling with high accuracy and good physical consistency. This invention can effectively learn historical scheduling rules and be used to simulate and predict the scheduling outflow process of reservoirs under long-term hydrological input.

[0006] To achieve the above objectives, this invention provides a rolling simulation method for reservoir scheduling that couples data-driven and physical constraints, comprising the following steps:

[0007] Step S1: Obtain the time-series operation data of the reservoir in historical periods. The time-series operation data includes reservoir water level, reservoir capacity, inflow, outflow and the number of days in the year for the corresponding period.

[0008] Step S2: Based on the time-series runtime data obtained in step S1, construct a set of feature variables and a set of target variables, wherein the set of feature variables includes the storage capacity at the beginning of the time period. Inbound flow and number of days in a year The target variable set is the storage capacity at the end of the time period. ;

[0009] Step S3: Using a machine learning algorithm, a reservoir scheduling simulation model is trained using the feature variable set and target variable set constructed in step S2. The reservoir scheduling simulation model learns the reservoir capacity from the feature variables to the end of the time period. Nonlinear mapping relationship: ;

[0010] Step S4: Use the reservoir scheduling simulation model trained in step S3 to perform rolling simulation prediction and generate a long-term reservoir capacity sequence.

[0011] Step S5: Based on the long series of reservoir capacity sequences generated in step S4, calculate the outflow for each time period using the water balance equation to obtain a preliminary outflow sequence; based on predefined reservoir operation constraints, verify and correct the preliminary outflow sequence, and output the final outflow sequence that satisfies the constraints.

[0012] Furthermore, the time scale of the time series data mentioned in step S1 is daily, and includes observation data for at least three consecutive years.

[0013] Furthermore, in step S3, a leave-out cross-validation method is used during model training to prevent overfitting. Specifically, 1 / 3 of the data is randomly selected for validation each time, and the remaining 2 / 3 of the data is used for model training. This cross-validation process is repeated 100 times to evaluate the stability of the model.

[0014] Furthermore, step S4 specifically includes:

[0015] S4.1: Initialize the initial reservoir capacity V (0) ;

[0016] S4.2: For the current time period t, the feature variables [ Input the reservoir scheduling simulation model to obtain the reservoir capacity at the end of the time period. ;

[0017] S4.3: The output of step S4.2 As the initial storage capacity for the next time period t+1 Combined with the inbound flow in the next period and number of days in a year Repeat step S4.2 to iteratively generate a long series of library capacity sequences { , , ..., }

[0018] Furthermore, in step S5, the reservoir operation constraints include at least one of the following: water balance constraints, reservoir water level constraints, reservoir discharge flow constraints, hydropower station output constraints, and non-negative constraints.

[0019] Furthermore, the water balance equation is as follows:

[0020] ;

[0021] Where Δt is the length of the calculation period.

[0022] A reservoir scheduling rolling simulation device that couples data-driven and physical-constrained methods includes:

[0023] The data acquisition module is used to acquire the time-series operation data of the reservoir in historical periods. The time-series operation data includes reservoir water level, reservoir capacity, inflow, outflow and the number of days in the year for the corresponding period.

[0024] The feature and target construction module is used to construct a feature variable set and a target variable set from the acquired time-series runtime data, wherein the feature variable set includes the initial storage capacity of the time period. Inbound flow and number of days in a year The target variable set is the storage capacity at the end of the time period. ;

[0025] The model training module is used to train a reservoir scheduling simulation model using machine learning algorithms on the constructed set of feature variables and the set of target variables. The reservoir scheduling simulation model learns the reservoir capacity from the feature variables to the end of the time period. Nonlinear mapping relationship: ;

[0026] The rolling prediction module is used to perform rolling simulation prediction using the trained reservoir scheduling simulation model to generate a long-term reservoir capacity sequence.

[0027] The flow calculation and constraint verification module is used to calculate the outflow for each time period based on the generated long series of reservoir capacity sequences and the water balance equation to obtain a preliminary outflow sequence; based on predefined reservoir operation constraints, the module verifies and corrects the preliminary outflow sequence and outputs the final outflow sequence that meets the constraints.

[0028] Furthermore, the time scale of the time series running data acquired by the data acquisition module is daily, and includes at least three consecutive years of observation data.

[0029] Furthermore, the model training module employs a leave-out cross-validation method during training to prevent overfitting. Specifically, 1 / 3 of the data is randomly selected for validation each time, and the remaining 2 / 3 of the data is used for model training. This cross-validation process is repeated 100 times to evaluate the stability of the model.

[0030] Furthermore, the rolling prediction module is specifically used for:

[0031] S4.1: Initialize the initial reservoir capacity V (0) ;

[0032] S4.2: For the current time period t, the feature variables [ Input the reservoir scheduling simulation model to obtain the reservoir capacity at the end of the time period. ;

[0033] S4.3: The output of step S4.2 As the initial storage capacity for the next time period t+1 Combined with the inbound flow in the next period and number of days in a year Repeat step S4.2 to iteratively generate a long series of library capacity sequences { , , ..., }

[0034] The advantages and beneficial effects of this invention are as follows:

[0035] (1) This invention ingeniously combines machine learning with industry knowledge and embeds multiple physical constraints, making the simulation results not only highly accurate, but also usable and reliable in actual engineering. The trained model can be used to simulate the reservoir scheduling process under different historical periods or future scenarios, and is suitable for applications such as assessing the impact of changes in hydrological conditions and simulating extreme flood events.

[0036] (2) Strict physical constraints and reliable simulation results. By “simulating the reservoir capacity first and then calculating the outflow”, the water balance constraints are strictly met. At the same time, engineering constraints such as water level, flow rate, head and output are embedded, and the simulation results meet the actual scheduling requirements.

[0037] (3) The scheduling is highly adaptable and accurate. By using “number of days in a year d” as a feature variable, the scheduling rules of different time periods can be distinguished. The machine learning model can efficiently learn complex nonlinear relationships, and the simulation results are highly consistent with the measured data.

[0038] (4) It has strong anti-overfitting ability. The model is trained by using the leave-blank method for cross-validation, which can still ensure generalization ability when the historical data period is short. Attached Figure Description

[0039] Figure 1 This is a flowchart of a reservoir scheduling rolling simulation method that couples data-driven and physical constraints, according to an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of rolling prediction according to an embodiment of the present invention;

[0041] Figure 3 This is a comparison of the outbound flow simulation process in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] like Figure 1 As shown, this embodiment of the invention provides a rolling simulation method for reservoir scheduling that couples data-driven and physical constraints, comprising the following steps:

[0044] 1) Data preparation: Obtain time-series operational data of the reservoir over historical periods, including reservoir water level, reservoir capacity, inflow, outflow, and the number of days in each period within the year; 2)

[0045] Feature and target construction: Initial storage capacity for each time period Inbound flow and number of days in a year As a model feature variable, the storage capacity at the end of the time period As the target variable of the model;

[0046] 3) Model Training: Using machine learning algorithms, the reservoir scheduling simulation model is trained with the features and targets constructed in step S2, learning the nonlinear mapping relationship f from features to reservoir capacity at the end of a time period, i.e. ;

[0047] 4) Rolling simulation prediction: Using a trained machine learning model, predictions are made time-by-time starting from an initial state; the prediction results for the current time period t are then used to generate predictions. As the initial storage capacity input for the next time period t+1, combined with the inflow rate for the next time period. and number of days in a year ,predict This process is repeated iteratively, generating a long series of changes in storage capacity;

[0048] 5) Outflow Calculation and Constraint Verification: Based on the reservoir capacity sequence obtained in step S4, the outflow is calculated using the water balance equation. Calculate the outflow rate for each time period; and during the calculation process, verify or correct the simulation results by combining predefined reservoir operation constraints.

[0049] Taking a certain reservoir as an example, the specific implementation steps are as follows:

[0050] Step S1, Data Preparation: Collect and obtain the daily reservoir water level, inflow and outflow data from 2012 to 2023 after the reservoir was fully put into operation.

[0051] Step S2, Feature and Target Construction: Based on the reservoir capacity curve, the reservoir water level is converted into reservoir capacity data, and the number of days in the year corresponding to each time period is extracted. To ensure that the simulation process strictly meets the water balance constraint (Equation 1), the model does not directly simulate the outflow, but instead simulates the reservoir capacity and then calculates the outflow through water balance (Equation 2). As shown in Equation 3, the reservoir capacity at the end of the time period is... The target variable is the inflow rate into the reservoir during that period. and the initial storage capacity of the period (or reservoir water level) As a characteristic variable, considering the differences in reservoir scheduling rules during different scheduling periods within the year, the number of days in the current year is used. It is also used as a feature variable.

[0052] (1);

[0053] (2);

[0054] (3);

[0055] In the formula: , The storage capacity (m) at the beginning and end of the time period are respectively. 3 ), for Inbound flow rate during the period (m 3 / s); for Reservoir discharge flow rate during the period (m³) 3 / s), including power generation flow and wastewater discharge flow; is the calculation period; f is the machine learning model; This represents the number of days in the current year.

[0056] Step S3, Model Training: In this example, a reservoir scheduling simulation model is established using the Gradient Boosting Machine (GBM) machine learning algorithm. The model is trained using the features and objectives constructed in Step S2 to learn the nonlinear mapping relationship f from features to the reservoir capacity at the end of a time period. Furthermore, to prevent overfitting, the LMOCV (leave-m-out cross-validation) method is used for rigorous cross-validation testing. Each time, 1 / 3 of the data is randomly selected for validation, and the remaining 2 / 3 is used for model training. This process is repeated 100 times to obtain the distribution of the cross-validation evaluation index.

[0057] GBM is a very popular machine learning algorithm, belonging to the Boosting algorithm family, and has been successfully and widely applied in many fields, such as classification and regression problems. This study focuses on regression problems. Its basic principle is to sequentially construct multiple weak learners and combine them into the existing model in an accumulative manner. The newly added weak learner is trained based on the negative gradient information of the current model's loss function. The goal is to reduce the cumulative model loss after adding the weak learner towards the negative gradient. Furthermore, the base learners are linearly combined with different weights, allowing high-performing learners to be reused. The most commonly used base learner is the decision tree.

[0058] The basic steps of GBM are as follows:

[0059] Given a sample sequence (x) i , y i Let y = 1, ..., n, where n is the number of samples. The goal is to fit a function F(x) such that the loss function L[y, F(x)] = [y - F(x)]. 2 Minimize / 2. That is, minimize the cumulative loss function J = ΣL[y, F(x)] by adjusting F(x). Treating F(x) as a parameter, the derivative is obtained as follows:

[0060] (4);

[0061] In the formula: x is the explanatory variable (independent variable), which in this study is a physical attribute, such as soil and topographic attributes; y is the response variable (dependent variable), which in this study is a hydrological model parameter; L is the loss function, which in this study uses the sum of squares loss; J is the cumulative loss function. F is the function to be fitted, that is, the function that constructs the relationship between the response variable and the explanatory variables.

[0062] Therefore, the residual yF(x) can be interpreted as the negative gradient g(x) of the loss function, i.e.:

[0063] (5);

[0064] GBM, borrowing the idea of ​​gradient descent, continuously updates and corrects the model based on the negative gradient direction of the current loss function in an iterative manner to search for the optimal function. To find the optimal function F(x), an initial value is set: F0(x) = h0(x). Treating the function F(x) as a whole, similar to the update process of gradient descent, it is assumed that the optimal function obtained after k iterations is:

[0065] (6);

[0066] Among them, h i (x) is

[0067] (7);

[0068] In the formula: h(x) is the weak learner; α is the learning rate, ranging from 0 to 1; g(x) is the gradient of the loss function.

[0069] When initializing the model, a simple model such as the average value Σy / n of the predicted output samples can be used. In the k-th iteration of gradient boosting, the algorithm does not modify the existing, potentially imperfect model F. k-1 Instead of using (x), a new model F is constructed by adding a weak learner h(x). k (x) = F k-1 The gradient boosting method uses h(x) + h(x) to improve the overall model accuracy. It assumes that the optimal learner h(x) should satisfy F(x) + h(x). k (x) = F k-1 (x) + h(x) = y, that is, h(x) = y - F k-1 (x). Therefore, GBM is the expression of h(x) and the residual y - F. k-1 (x) is fitted to obtain F. k To continuously correct F k-1 Furthermore, the residual is in the direction of the negative gradient of the loss function.

[0070] Steps S4-S5, rolling simulation and outbound flow calculation: (e.g.) Figure 2 As shown, using a trained GBM model, predictions are made time-by-time starting from an initial state; the prediction results for the current time period t are then used. As the next period The initial storage capacity input, combined with the inbound flow rate for the next period. and number of days in a year ,predict This process is repeated iteratively to generate a long series of reservoir capacity changes. At the same time, the reservoir discharge is calculated according to the water balance formula for each time period, and the simulation results are verified or corrected in combination with predefined reservoir operation constraints to obtain a long series of daily discharge simulation processes for a certain reservoir from 2012 to 2023.

[0071] The specific constraints on reservoir operation include:

[0072] 1) Water balance constraint

[0073] (8);

[0074] In the formula: To calculate unit duration, and Reservoirs at Inbound and outbound flow values ​​for a given period For the reservoir in The storage capacity value at any given time.

[0075] 2) Reservoir water level constraints

[0076] (9);

[0077] In the formula: for The upstream water level at any given time, , They are respectively The minimum and maximum water levels that are allowed at any given time.

[0078] 3) Reservoir discharge constraints

[0079] (10);

[0080] In the formula: , The upper and lower limits of outbound flow for a given period. This is a function relating upstream water level to maximum discharge capacity.

[0081] 4) Power output constraints of hydropower stations

[0082] (11);

[0083] In the formula, For time period The water purifier head, It is a function relating the head of the hydropower station to the expected output. , The lower and upper limits of power output for a given period are generally determined by a combination of factors, including the power plant's installed capacity, the rated output of the generating units, the vibration zone, and peak-shaving requirements.

[0084] 5) Nonnegativity constraint

[0085] All variables must be non-negative.

[0086] The accuracy of the model is evaluated using indicators such as Pearson correlation coefficient r, Nash efficiency coefficient NSE, Kling-Gupta efficiency coefficient KGE, root mean square error RMSE, and mean absolute error MAE.

[0087] (12);

[0089] (13);

[0091] (14);

[0092]

[0093] (15);

[0095] (16);

[0097] In the formula: X o X s These are the measured sequence and the simulated sequence, respectively. , , respectively, are the mean values ​​of the measured sequence and the simulated sequence; n is the sequence length.

[0098] Figure 3 The study compares the simulated outflow process of the GBM model with the actual outflow process. From a daily scale perspective, the simulated outflow process and the actual outflow process show good consistency, with the scatter points basically distributed along a 1:1 diagonal. The evaluation indices are r=0.96, NSE=0.92, KGE=0.96, RMSE=2224, and MAE=1547, respectively. Overall, the data-driven and physically constrained reservoir rolling simulation method in this embodiment has good simulation results and can be effectively applied to the scheduling simulation of a certain reservoir, providing technical support for reservoir scheduling and adaptive management under changing environments.

[0099] This invention also provides a reservoir scheduling rolling simulation device that couples data-driven and physical constraints, comprising:

[0100] The data acquisition module is used to acquire the time-series operation data of the reservoir in historical periods. The time-series operation data includes reservoir water level, reservoir capacity, inflow, outflow and the number of days in the year for the corresponding period.

[0101] The feature and target construction module is used to construct a feature variable set and a target variable set from the acquired time-series runtime data, wherein the feature variable set includes the initial storage capacity of the time period. Inbound flow and number of days in a year The target variable set is the storage capacity at the end of the time period. ;

[0102] The model training module is used to train a reservoir scheduling simulation model using machine learning algorithms on the constructed set of feature variables and the set of target variables. The reservoir scheduling simulation model learns the reservoir capacity from the feature variables to the end of the time period. Nonlinear mapping relationship: ;

[0103] The rolling prediction module is used to perform rolling simulation prediction using the trained reservoir scheduling simulation model to generate a long-term reservoir capacity sequence.

[0104] The flow calculation and constraint verification module is used to calculate the outflow for each time period based on the generated long series of reservoir capacity sequences and the water balance equation to obtain a preliminary outflow sequence; based on predefined reservoir operation constraints, the module verifies and corrects the preliminary outflow sequence and outputs the final outflow sequence that meets the constraints.

[0105] This invention has the following characteristics:

[0106] 1. High precision and strong physical consistency: This invention deeply couples data-driven machine learning with physical constraints. By predicting reservoir capacity and then back-calculating outflow, it strictly ensures water balance from the source of modeling. It also verifies the results by combining multiple engineering constraints such as water level, flow rate, and output. This ensures that the simulation results are not only highly accurate, but also fully comply with the physical rules and safety requirements of actual reservoir operation, thus solving the problem that traditional pure data-driven models are physically infeasible.

[0107] 2. Strong adaptability and generalization ability: The model uses initial reservoir capacity, inflow, and number of days in a year as core features to effectively capture the rule differences in reservoir scheduling at different times (such as flood season and non-flood season). The rolling prediction mechanism and rigorous cross-validation strategy ensure the model's stability in long-term simulations, enabling it to adapt to changing hydrological conditions and possessing good generalization ability, making it suitable for historical review and future scenario simulation.

[0108] 3. High engineering applicability and decision support value: The system provides a complete process from data processing and model training to rolling simulation and constraint verification, with directly usable output results. This method provides efficient and reliable core technical tools for intelligent reservoir scheduling, extreme event early warning, and scheduling scheme evaluation, significantly improving the intelligence and scientific nature of scheduling decisions.

[0109] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A rolling simulation method for reservoir scheduling that couples data-driven and physical constraints, characterized in that, Includes the following steps: Step S1: Obtain the time-series operation data of the reservoir in historical periods. The time-series operation data includes reservoir water level, reservoir capacity, inflow, outflow and the number of days in the year for the corresponding period. Step S2: Based on the time-series runtime data obtained in step S1, construct a set of feature variables and a set of target variables, wherein the set of feature variables includes the storage capacity at the beginning of the time period. Inbound flow and number of days in a year The target variable set is the storage capacity at the end of the time period. ; Step S3: Using a machine learning algorithm, a reservoir scheduling simulation model is trained using the feature variable set and target variable set constructed in step S2. The reservoir scheduling simulation model learns the reservoir capacity from the feature variables to the end of the time period. Nonlinear mapping relationship: ; Step S4: Use the reservoir scheduling simulation model trained in step S3 to perform rolling simulation prediction and generate a long-term reservoir capacity sequence. Step S5: Based on the long series of reservoir capacity sequences generated in step S4, calculate the outflow for each time period using the water balance equation to obtain a preliminary outflow sequence; based on predefined reservoir operation constraints, verify and correct the preliminary outflow sequence, and output the final outflow sequence that satisfies the constraints.

2. The method according to claim 1, characterized in that, The time scale of the time series data mentioned in step S1 is daily, and includes observation data for at least 3 consecutive years.

3. The method according to claim 1, characterized in that, In step S3, a leave-out cross-validation method is used during model training to prevent overfitting. Specifically, 1 / 3 of the data is randomly selected for validation each time, and the remaining 2 / 3 of the data is used for model training. This cross-validation process is repeated 100 times to evaluate the stability of the model.

4. The method according to claim 1, characterized in that, Step S4 specifically includes: S4.1: Initialize the initial reservoir capacity V (0) ; S4.2: For the current time period t, the feature variables [ Input the reservoir scheduling simulation model to obtain the reservoir capacity at the end of the time period. ; S4.3: The output of step S4.2 As the initial storage capacity for the next time period t+1 Combined with the inbound flow in the next period and number of days in a year Repeat step S4.2 to iteratively generate a long series of library capacity sequences { , , ..., } 5. The method according to claim 1, characterized in that, In step S5, the reservoir operation constraints include at least one of the following: water balance constraints, reservoir water level constraints, reservoir discharge flow constraints, hydropower station output constraints, and non-negative constraints.

6. The method according to claim 1, characterized in that, The water balance equation is as follows: ； Where Δt is the length of the calculation period.

7. A reservoir scheduling rolling simulation device coupling data-driven and physical constraint, characterized in that, include: The data acquisition module is used to acquire the time-series operation data of the reservoir in historical periods. The time-series operation data includes reservoir water level, reservoir capacity, inflow, outflow and the number of days in the year for the corresponding period. The feature and target construction module is used to construct a feature variable set and a target variable set from the acquired time-series runtime data, wherein the feature variable set includes the initial storage capacity of the time period. Inbound flow and number of days in a year The target variable set is the storage capacity at the end of the time period. ; The model training module is used to train a reservoir scheduling simulation model using machine learning algorithms on the constructed set of feature variables and the set of target variables. The reservoir scheduling simulation model learns the reservoir capacity from the feature variables to the end of the time period. Nonlinear mapping relationship: ; The rolling prediction module is used to perform rolling simulation prediction using the trained reservoir scheduling simulation model to generate a long-term reservoir capacity sequence. The flow calculation and constraint verification module is used to calculate the outflow for each time period based on the generated long series of reservoir capacity sequences and the water balance equation to obtain a preliminary outflow sequence; based on predefined reservoir operation constraints, the module verifies and corrects the preliminary outflow sequence and outputs the final outflow sequence that meets the constraints.

8. The apparatus according to claim 7, characterized in that, The time scale of the time series operation data acquired by the data acquisition module is daily, and includes observation data for at least 3 consecutive years.

9. The apparatus according to claim 7, characterized in that, The model training module employs a leave-out cross-validation method during training to prevent overfitting. Specifically, 1 / 3 of the data is randomly selected for validation each time, and the remaining 2 / 3 of the data is used for model training. This cross-validation process is repeated 100 times to evaluate the stability of the model.

10. The apparatus according to claim 7, characterized in that, The rolling prediction module is specifically used for: S4.1: Initialize the initial reservoir capacity V (0) ; S4.2: For the current time period t, the feature variables [ Input the reservoir scheduling simulation model to obtain the reservoir capacity at the end of the time period. ; S4.3: The output of step S4.2 As the initial storage capacity for the next time period t+1 Combined with the inbound flow in the next period and number of days in a year Repeat step S4.2 to iteratively generate a long series of library capacity sequences { , , ..., }

Images