Pumped storage power station water level, water head and reservoir capacity prediction method considering time factor
By combining support vector machine and XGBoost algorithm, and taking into account time factors and the number of pumping points for power generation, a model of the correspondence between water level and reservoir capacity is generated. This solves the problem of inaccurate prediction in existing technologies, achieves high-precision prediction of water level, reservoir capacity and head, verifies the rationality of the scheduling plan, and ensures the safety of the power plant.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGDIAN NEW ENERGY CO LTD
- Filing Date
- 2025-12-03
- Publication Date
- 2026-05-08
AI Technical Summary
In existing pumped storage power stations, the forecasting methods for power generation and pumping plans fail to effectively consider time factors and environmental impacts, resulting in inaccurate forecasts and an inability to accurately assess the rationality of the plans. This may lead to water levels being lower or higher than the adjustable water level, affecting unit power generation or pumping, and even exacerbating problems on the reservoir bank and dam.
A support vector machine is used to generate a model of the correspondence between water level and reservoir capacity. Combined with the XGBoost algorithm, the water level changes in the upper and lower reservoirs are predicted by time information and the number of power generation and pumping points. Bayesian optimization is used to find the optimal hyperparameters and to construct a model of the relationship between power generation, the number of pumping points and water level changes. The water head and reservoir capacity are predicted by combining the water level-reservoir capacity relationship.
It improves prediction accuracy, reduces errors, accurately verifies the rationality of scheduling plans, adapts to engineering realities, reduces training costs, and achieves multi-objective integrated prediction of water level, reservoir capacity, and head, thus ensuring the safety of power plants.
Smart Images
Figure CN121998153A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of pumped storage power station technology, and in particular to a method for predicting water level, head, and reservoir capacity of pumped storage power stations that takes into account time factors. Background Technology
[0002] The daily power generation and pumping plans for pumped-storage power stations are issued by the grid dispatch center the day before. Power generation and pumping operations cause changes in water levels and heads in the upper and lower reservoirs. For the upper and lower reservoirs of a pumped-storage power station, these can be categorized as dead capacity and regulating capacity, with corresponding dead water levels and adjustable water levels. If the power generation and pumping plans are inappropriate, resulting in water levels below or above the adjustable water level, it will not only affect the normal power generation or pumping operations but also exacerbate stress changes on the reservoir banks and dam body, potentially leading to cracks and leakage in the long term. Therefore, it is essential to predict changes in water levels and heads in the upper and lower reservoirs in advance using the pumping and power generation plans issued by the grid dispatch center, and to verify the rationality of the power generation and pumping plans.
[0003] Currently, the mainstream forecasting methods rely on experimentally derived formulas relating the number of power generation and pumping stations to changes in reservoir capacity. These formulas are used to calculate the capacity changes in the upper and lower reservoirs, and then the changes in water level are derived from the correlation between reservoir capacity and water level to predict the head. However, in practical applications, the correlation between the number of power generation and pumping stations and changes in reservoir capacity is not constant. Using a single mathematical formula can easily lead to errors. Furthermore, this method does not consider the influence of evaporation, precipitation, and other environmental factors, resulting in inaccurate predictions that cannot be used as a criterion for judging the rationality of power generation and pumping plans. Summary of the Invention
[0004] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a method for predicting water level, head and reservoir capacity of pumped storage power stations that takes into account time factors, so as to solve the problems mentioned in the background art.
[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for predicting water level, head, and reservoir capacity of pumped storage power stations considering time factors, comprising the following steps: Step 1: Obtain the corresponding data of water level and reservoir capacity during the water storage period of the upper and lower reservoirs of the pumped storage power station. Use a support vector machine to fit the corresponding data to generate water level-reservoir capacity curves of the upper and lower reservoirs, and obtain the corresponding relationship model of water level and reservoir capacity of the upper and lower reservoirs. Step 2: Determine the input variables of the XGBoost model and perform preprocessing. The input variables include time information, number of pumping and power generation points, pumping / power generation mode identifier, and initial upper and lower reservoir water levels. Preprocessing includes time information encoding, normalization of the number of pumping and power generation points, and normalization of the initial upper and lower reservoir water levels to obtain standardized input data. Step 3: Randomly extract a portion of the data from the standardized input data as the Bayesian optimization dataset, and use the Bayesian optimization algorithm to optimize the hyperparameters of the XGBoost model to determine the optimal hyperparameter combination; Step 4: Divide the standardized input data into a training set, a validation set, and a test set. Based on the optimal hyperparameter combination, train the XGBoost model using the training set and the validation set to obtain the relationship model between the number of power generation / pumping points and the change in water level between the upper and lower reservoirs. Verify the model performance using the test set. Step 5: Input the number of pumping and power generation points, target time information, target pumping / power generation mode, and target initial upper and lower reservoir water levels in the target pumping / power generation plan. Obtain the change in upper and lower reservoir water levels using the XGBoost model trained in Step 4. Combine the target initial upper and lower reservoir water levels to obtain the predicted upper and lower reservoir water levels. Using the water level-storage capacity correspondence model obtained in Step 1, obtain the predicted upper and lower reservoir capacity based on the predicted upper and lower reservoir water levels. Subtract the predicted upper and lower reservoir water levels to obtain the predicted head.
[0006] Preferably, in step 1, the support vector machine (SVM) adopts the support vector regression (SVR) model, and the optimal parameters of the SVM are determined by a grid search algorithm. The optimal parameters include the regularization parameter C and the kernel function.
[0007] Preferably, the kernel function is selected from one of the following: RBF kernel function, linear kernel function, polynomial kernel function, or neural network kernel function.
[0008] Preferably, the regularization parameter C has a value range of [0.001, 1000].
[0009] Preferably, in step 1, the water level is used as the input parameter of the SVM, and the reservoir capacity is used as the output parameter of the SVM. The SVM model is trained using the corresponding data of water level and reservoir capacity in the early stage of water storage in the upper and lower reservoirs to obtain the corresponding relationship model of water level and reservoir capacity in the upper and lower reservoirs.
[0010] Preferably, in step 2, the time information includes month information, day information, and minute information. The time information is encoded using a sine / cosine encoding method, and the encoding formula is as follows: Sin_Mon=sin(2π×Month / 12), Cos_Mon=cos(2π×Month / 12); Sin_Day=sin(2π×Day / 31), Cos_Day=cos(2π×Day / 31); Sin_Min=sin(2π×Minute / 1440), Cos_Min=cos(2π×Minute / 1440); Where Month represents the month number, Day represents the day of the month, and Minute represents the minute of the day. The encoded time information is in the following format: Time=[(Sin_Mon,Cos_Mon),(Sin_Day,Cos_Day),(Sin_Min,Cos_Min)].
[0011] Preferably, in step 2, the normalization formula for the number of pumped power generation points is as follows: Regarding the number of power generation points: D f =D1 / (a×96); Regarding the number of pumping points: D f =D2 / (a×96); Where D1 represents the actual number of power generation points issued, D2 represents the actual number of pumping points issued, a represents the number of units in the pumped storage power station, 96 represents the maximum number of points per unit per day, and D... f This represents the normalized number of pumped-hydro power generation points.
[0012] Preferably, in step 2, the pumping / power generation mode identifier is defined as follows: pumping mode corresponds to identifier 0, and power generation mode corresponds to identifier 1; the initial upper and lower reservoir water levels, after normalization, are represented as H=[H s H x ], where H s H represents the normalized initial water level of the upper reservoir. x This represents the normalized initial water level of the lower reservoir.
[0013] Preferably, in step 3, the objective function f(p) of the Bayesian optimization is defined as:
[0014] Where p is the hyperparameter space, n is the number of test sets in the Bayesian optimization dataset, e(p,i) is the predicted value of the i-th data in the test set under this hyperparameter combination, real is the true value of the i-th data in the test set, and the smaller the objective function value, the better the hyperparameter optimization effect.
[0015] Preferably, the Bayesian optimization employs a GP surrogate model, which selects the optimal sampling point by maximizing the expected increment EI method and iteratively updates the surrogate model to obtain the globally optimal hyperparameters.
[0016] Preferably, in step 3, the hyperparameters of the XGBoost model include the learning rate, the number of decision trees (n_estimators), the maximum tree depth (depth_max), and the minimum number of nodes (min_child).
[0017] Preferably, in step 4, the training set is used for model parameter fitting, the validation set is used to monitor overfitting during model training and adjust parameters, the test set is used to evaluate the prediction accuracy of the trained model, and the extraction ratio of the Bayesian optimization dataset is determined according to the data size to ensure the reliability of hyperparameter optimization.
[0018] Preferably, in step 5, the change in water level between the upper and lower reservoirs is directly output through the XGBoost model, and the predicted water level between the upper and lower reservoirs = the target initial water level between the upper and lower reservoirs + the corresponding change in water level; the predicted reservoir capacity between the upper and lower reservoirs is calculated by inputting the predicted water level between the upper and lower reservoirs into the water level-capacity correspondence model obtained in step 1; the predicted head is the predicted water level of the upper reservoir - the predicted water level of the lower reservoir.
[0019] Beneficial effects of this invention: 1. This invention uses support vector machines to generate a model of the correspondence between reservoir water level and reservoir capacity, which can effectively avoid errors caused by other influencing factors. The XGBoost algorithm with Bayesian optimization is used to predict the water level changes of the upper and lower reservoirs through time information and the number of power generation pumping points, thereby obtaining water head and corresponding reservoir capacity information.
[0020] 2. Higher prediction accuracy: This invention uses SVM to accurately fit the nonlinear relationship between water level and reservoir capacity, and combines the Bayesian-optimized XGBoost model to explore the correlation of water level changes, thereby reducing fitting and prediction errors. The results can be used as a criterion for the rationality of scheduling plans.
[0021] 3. Strong scenario adaptability: This invention incorporates sine and cosine encoding of time information, adapting to the seasonal characteristics of evaporation and precipitation, as well as the 15-minute interval scheduling requirements of power plants, and the input data processing is in line with the actual engineering situation.
[0022] 4. Training efficiency optimization: The present invention uses Bayesian optimization to efficiently obtain the optimal hyperparameters of XGBoost, and combined with the parallel computing capability of the algorithm itself, it reduces training overhead.
[0023] 5. Multi-objective integrated prediction: This invention combines SVM and XGBoost to output the prediction results of water level, reservoir capacity and head in one go, without the need for secondary calculation.
[0024] 6. This invention solves the problems of large errors and failure to take into account the influence of environment and time in existing prediction methods. It can accurately capture seasonal and scheduling cycle characteristics, with high prediction accuracy and excellent training efficiency. It can output water level, reservoir capacity and head results in an integrated manner, effectively verify the rationality of scheduling plans, ensure power plant safety, and is easy to implement in engineering projects. Attached Figure Description
[0025] Figure 1 A flowchart of a method for predicting water level, head, and reservoir capacity of a pumped storage power station that takes time factors into account; Figure 2 Flowchart for the XGBoost algorithm implementation; Figure 3 This is a diagram illustrating the training process of XGboost. Detailed Implementation
[0026] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0027] Example 1: A method for predicting water level, head, and reservoir capacity of a pumped storage power station. First, based on the corresponding data of water level and reservoir capacity during the impoundment period of the upper and lower reservoirs of the pumped storage power station, a model is fitted. Because the relationship between water level and reservoir capacity is non-linear, polynomial regression and traditional interpolation algorithms are prone to large errors. Therefore, a Support Vector Machine (SVM) is used to fit the data, generating upper and lower reservoir water level-capacity curves, thus obtaining a model of the correspondence between upper and lower reservoir water levels and capacity. Then, the XGBoost algorithm is used to construct a model of the relationship between the number of power generation and pumping points and the changes in upper and lower reservoir water levels. Considering that the main influencing factors of reservoir water level, evaporation and precipitation, have strong seasonal factors, time information is used as the input variable of the model. After obtaining the relationship between the number of power generation and pumping points and the changes in upper and lower reservoir water levels, the water levels after power generation or pumping can be obtained from the initial upper and lower reservoir water levels. Subtracting the upper and lower reservoir water levels yields the head. Using the water level-capacity correspondence model, the predicted reservoir capacity of the upper and lower reservoirs can be obtained. The flowchart of this invention is as follows: Figure 1 As shown.
[0028] The specific working process and principle are as follows: First, a model relating water level and reservoir capacity needs to be obtained. Essentially, this is a regression problem, and a Support Vector Machine (SVM) Support Vector Regression (SVR) model is used. SVM is a supervised learning model that uses a "kernel trick" to map the original nonlinear problem to a high-dimensional feature space, making it linearly separable from the original nonlinear problem, thus solving complex nonlinear classification tasks. SVM requires determining two important parameters: the regularization parameter C and the kernel function selection. Generally, C ranges from [0.001 to 1000]. Common kernel function types include RBF kernel, linear kernel, polynomial kernel, and neural network kernel. Because the number of hyperparameters required is relatively small, a grid search is used to obtain the optimal parameters for SVM. A grid search exhaustively searches for all possible hyperparameters to find the optimal hyperparameter space. Water level is used as the input parameter of SVM, and reservoir capacity as the output parameter. The model is trained using the initial water level and capacity data of the upper and lower reservoirs to obtain the corresponding relationship model between the water level and capacity of the upper and lower reservoirs.
[0029] Then, the XGBoost algorithm was used to construct a model of the relationship between the number of power generation and pumping points and the water levels of the upper and lower reservoirs. XGBoost (eXtreme Gradient Boosting) is an efficient gradient boosting decision tree (GBDT) framework, proposed by Chen Tianqi's team in 2016. By integrating multiple weakly learned CART trees (decision trees) and combining gradient boosting algorithms with engineering optimization techniques, it has become a mainstream tool for structured data modeling. Its core advantages include: high performance and efficiency, utilizing pre-sorted feature block storage (Block structure) to support multi-threaded parallel computing, combined with cache-aware access and sparse data processing (automatically learning the splitting direction for missing values), significantly reducing memory and computational overhead; high prediction accuracy, employing a second-order Taylor expansion to approximate the objective function, while combining first-order and second-order gradients (Hessian matrix) to optimize the splitting strategy, resulting in higher accuracy compared to traditional GBDT which only uses the first-order gradient; and the introduction of L1 / L2 regularization terms (through parameters lambda and alpha) and tree structure complexity penalty (parameter gamma) to effectively suppress overfitting; strong flexibility and scalability, supporting custom objective functions and evaluation metrics (such as weighted loss in financial risk control), suitable for various tasks such as classification, regression, and ranking (such as recommendation systems). It is also compatible with multi-language interfaces such as Python / R / Java and can be exported to ONNX and other formats for deployment in production environments. With these features, XGBoost is widely used in Kaggle competitions and industrial scenarios, especially excelling in high-dimensional data and nonlinear relationship modeling. Taking all the above factors into consideration, this patent chooses to use XGBoost as the method for constructing the regression prediction model.
[0030] The XGBoost model combines multiple weak learners (i.e., decision trees) to build a powerful prediction model. It is applied to construct a model relating resource allocation and production cycle time. The basic idea is as follows: 1) Initialize the predicted value for each input sample data, define the loss function, and obtain the first decision tree that fits the sample data; 2) Calculate the first and second derivatives of the loss function with respect to each sample prediction value; 3) Use the fitting error as the training objective for the second decision tree, and build a new decision tree based on the calculated derivative information; 4) Use the new decision tree to predict sample values and add them to the original values; 5) Continue in this manner. If the stopping condition is met, finally, the prediction results of multiple decision trees are weighted and summed to obtain a more accurate overall model. The XGBoost algorithm implementation flowchart is shown below. Figure 2 As shown; The training process of XGboost is as follows: Figure 3 As shown; First, the input data for the model is determined, which includes four variables: time information, number of pumping and power generation points, selection of pumping or power generation, and initial upper and lower reservoir water levels. Considering the periodicity of time information, sine and cosine are used to encode the time information. For pumped storage power stations, taking the East China Power Grid as an example, from the moment the unit starts pumping or power generation, the commands issued by the grid dispatch center are generally at 15-minute intervals. Therefore, the time encoding needs to consider the minute level. Thus, the time information in this patent mainly includes month information, day information, and minute information. The format of the time information is: Time=[(Sin_Mon,Cos_Mon),(Sin_Day,Cos_Day),(Sin_Min,Cos_Min)], as shown in the formula below: Sin_Mon=sin(2π×Month / 12), Cos_Mon=cos(2π×Month / 12); Sin_Day=sin(2π×Day / 31), Cos_Day=cos(2π×Day / 31); Sin_Min=sin(2π×Minute / 1440), Cos_Min=cos(2π×Minute / 1440); In the formula: Month Indicates which month, Day Indicates the day. Minute Indicates the minute of the day; The data for pumped-storage power generation points needs to be normalized. One power generation point represents 15 minutes of operation at rated power under power generation conditions. The same applies to pumping points. A single unit can generate a maximum of 96 points per day. If a pumped-storage power station has *a* units, the theoretical total number of points per day is *a* *96. The same applies to pumping points. The normalization process is as follows:
[0031] In the formula: D1 represents the actual number of power generation points issued during this period, D f D represents the normalized number of power generation points, and D2 represents the actual number of pumping points issued during this period. f This represents the normalized number of pumping points.
[0032] The power generation or pumping mode is selected; 0 indicates pumping, and 1 indicates power generation. The initial water levels of the upper and lower reservoirs are represented by H=[H]. s H x ], H s H represents the normalized initial water level of the upper reservoir. x This represents the normalized initial water level of the lower reservoir. All input data has been processed. Date=[Time,D f (D c),1(0),H].
[0033] After determining the dataset, a portion is randomly selected from it as the dataset for Bayesian optimization. The Bayesian optimization algorithm is then used to optimize the hyperparameters of XGBoost, primarily including: learning rate, number of decision trees (n_estimators), maximum tree depth (depth_max), and minimum number of nodes (min_child). Once the optimal hyperparameters for XGBoost are set, model training can begin.
[0034] The steps of Bayesian optimization are as follows: First, we define an objective function, primarily considering mean average precision (mAP) and inference efficiency. Assume the given hyperparameter space is... p objective function f ( p This can be represented as:
[0035] n This indicates the number of test sets in the Bayesian optimization dataset. e ( p , i ) indicates the first (instance) in the test set under this set of hyperparameters. i Predicted values of the data points real Indicates the first test set i The true value of each data point, at this time f ( p This can be expressed as the sum of errors of this set of hyperparameters, which is useful for prediction. f ( p The smaller the value of ), the better the performance optimization of the model is by this set of hyperparameters.
[0036] Subsequently, Bayesian optimization uses a selected GP surrogate model to approximate the behavior of the objective function. The surrogate model is then progressively built by selecting a series of sampling points, balancing exploration, unexplored regions, and the utilization of high-value regions. In each iteration, the surrogate model selects an optimal sampling point based on the expected improvement (EI) method. This point is then evaluated against the true objective function to obtain the evaluation result.
[0037] Finally, the parameters and prediction results of the surrogate model are updated to more accurately estimate the objective function. Bayesian optimization iteratively optimizes the surrogate model step by step to find the global optimum. After obtaining the optimal hyperparameter space, the model ontology can be trained.
[0038] The training of the model involves dividing the dataset into a training set, a validation set, and a test set. The training set and validation set are used to train the model, and the test set is used to test and evaluate the performance of the trained model.
[0039] It is recommended to use Python for the implementation and training of SVM and XGBoost models, C# for deployment and application, and WinForms for the UI.
[0040] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.
Claims
1. A method for predicting water level, head, and reservoir capacity of a pumped-storage power station considering time factors, characterized in that, Includes the following steps: Step 1: Obtain the corresponding data of water level and reservoir capacity during the water storage period of the upper and lower reservoirs of the pumped storage power station. Use a support vector machine to fit the corresponding data to generate water level-reservoir capacity curves of the upper and lower reservoirs, and obtain the corresponding relationship model of water level and reservoir capacity of the upper and lower reservoirs. Step 2: Determine the input variables of the XGBoost model and perform preprocessing. The input variables include time information, number of pumping and power generation points, pumping / power generation mode identifier, and initial upper and lower reservoir water levels. Preprocessing includes time information encoding, normalization of the number of pumping and power generation points, and normalization of the initial upper and lower reservoir water levels to obtain standardized input data. Step 3: Randomly extract a portion of the data from the standardized input data as the Bayesian optimization dataset, and use the Bayesian optimization algorithm to optimize the hyperparameters of the XGBoost model to determine the optimal hyperparameter combination. Step 4: Divide the standardized input data into a training set, a validation set, and a test set. Based on the optimal hyperparameter combination, train the XGBoost model using the training set and the validation set to obtain the relationship model between the number of power generation / pumping points and the change in water level between the upper and lower reservoirs. Verify the model performance using the test set. Step 5: Input the number of pumping and power generation points, target time information, target pumping / power generation mode, and target initial upper and lower reservoir water levels in the target pumping / power generation plan. Obtain the change in upper and lower reservoir water levels using the XGBoost model trained in Step 4. Combine the target initial upper and lower reservoir water levels to obtain the predicted upper and lower reservoir water levels. Using the water level-storage capacity correspondence model obtained in Step 1, obtain the predicted upper and lower reservoir capacity based on the predicted upper and lower reservoir water levels. Subtract the predicted upper and lower reservoir water levels to obtain the predicted head.
2. The method for predicting water level, head, and reservoir capacity of a pumped-storage power station considering time factors, as described in claim 1, is characterized in that... In step 1, the Support Vector Machine (SVM) adopts the Support Vector Regression (SVR) model, and the optimal parameters of the SVM are determined by the grid search algorithm. The optimal parameters include the regularization parameter C and the kernel function.
3. The method for predicting water level, head, and reservoir capacity of a pumped storage power station considering time factors, as described in claim 2, is characterized in that... The kernel function is selected from one of the following: RBF kernel function, linear kernel function, polynomial kernel function, or neural network kernel function.
4. The method for predicting water level, head, and reservoir capacity of a pumped-storage power station considering time factors, as described in claim 2, is characterized in that... The regularization parameter C has a value range of [0.001, 1000].
5. The method for predicting water level, head, and reservoir capacity of a pumped-storage power station considering time factors according to claim 1, characterized in that, In step 1, the water level is used as the input parameter of the SVM and the reservoir capacity is used as the output parameter of the SVM. The SVM model is trained using the corresponding data of water level and reservoir capacity in the early stage of water storage in the upper and lower reservoirs to obtain the corresponding relationship model of water level and reservoir capacity in the upper and lower reservoirs.
6. The method for predicting water level, head, and reservoir capacity of a pumped-storage power station considering time factors according to claim 1, characterized in that, In step 2, the time information includes month information, day information, and minute information. The time information is encoded using a sine / cosine encoding method, and the encoding formula is as follows: Sin_Mon=sin(2π×Month / 12), Cos_Mon=cos(2π×Month / 12); Sin_Day=sin(2π×Day / 31), Cos_Day=cos(2π×Day / 31); Sin_Min=sin(2π×Minute / 1440), Cos_Min=cos(2π×Minute / 1440); Where Month represents the month number, Day represents the day of the month, and Minute represents the minute of the day. The encoded time information is in the following format: Time=[(Sin_Mon,Cos_Mon),(Sin_Day,Cos_Day),(Sin_Min,Cos_Min)].
7. The method for predicting water level, head, and reservoir capacity of a pumped storage power station considering time factors according to claim 1, characterized in that, In step 2, the normalization formula for the number of pumped power generation points is as follows: Regarding the number of power generation points: D f =D1 / (a×96); Regarding the number of pumping points: D f =D2 / (a×96); Where D1 represents the actual number of power generation points issued, D2 represents the actual number of pumping points issued, a represents the number of units in the pumped storage power station, 96 represents the maximum number of points per unit per day, and D... f This represents the normalized number of pumped-hydro power generation points.
8. The method for predicting water level, head, and reservoir capacity of a pumped-storage power station considering time factors according to claim 1, characterized in that, In step 2, the pumping / power generation mode identifier is defined as follows: pumping mode corresponds to identifier 0, and power generation mode corresponds to identifier 1; the initial upper and lower reservoir water levels, after normalization, are represented as H=[H s H x ], where H s H represents the normalized initial water level of the upper reservoir. x This represents the normalized initial water level of the lower reservoir.
9. The method for predicting water level, head, and reservoir capacity of a pumped storage power station considering time factors according to claim 1, characterized in that, In step 3, the objective function f(p) of the Bayesian optimization is defined as: Where p is the hyperparameter space, n is the number of test sets in the Bayesian optimization dataset, e(p,i) is the predicted value of the i-th data in the test set under this hyperparameter combination, real is the true value of the i-th data in the test set, and the smaller the objective function value, the better the hyperparameter optimization effect.
10. The method for predicting water level, head, and reservoir capacity of a pumped storage power station considering time factors according to claim 1, characterized in that, The Bayesian optimization employs a GP surrogate model, which selects the optimal sampling point by maximizing the expected increment EI method and iteratively updates the surrogate model to obtain the globally optimal hyperparameters.
11. The method for predicting water level, head, and reservoir capacity of a pumped storage power station considering time factors according to claim 1, characterized in that, In step 3, the hyperparameters of the XGBoost model include the learning rate, the number of decision trees (n_estimators), the maximum tree depth (depth_max), and the minimum number of nodes (min_child).
12. The method for predicting water level, head, and reservoir capacity of a pumped-storage power station considering time factors according to claim 1, characterized in that, In step 4, the training set is used for model parameter fitting, the validation set is used to monitor overfitting during model training and adjust parameters, the test set is used to evaluate the prediction accuracy of the trained model, and the extraction ratio of the Bayesian optimization dataset is determined according to the data size to ensure the reliability of hyperparameter optimization.
13. The method for predicting water level, head, and reservoir capacity of a pumped storage power station considering time factors according to claim 1, characterized in that, In step 5, the change in water level between the upper and lower reservoirs is directly output through the XGBoost model, and the predicted water level between the upper and lower reservoirs = the target initial water level between the upper and lower reservoirs + the corresponding change in water level; the predicted reservoir capacity between the upper and lower reservoirs is calculated by inputting the predicted water level between the upper and lower reservoirs into the water level-capacity correspondence model obtained in step 1; the predicted head = predicted water level of the upper reservoir - predicted water level of the lower reservoir.