Radial flow type hydropower station water level prediction method based on hypothesis data enhancement and virtual parameters

Abstract

The invention discloses a runoff hydropower station water level prediction method based on hypothesis data enhancement and virtual parameters, which comprises the following steps: screening recent measured data and similar scheduling process data, and setting a boundary; establishing a total storage flow proportional relation ignoring intra-day distribution, and solving intra-day generalized storage flow; establishing a physical generalization model based on the virtual water consumption rate, and proposing a virtual water consumption rate limiting rule; multiple assumptions are proposed, and according to an assumption rule and a virtual water consumption rate model, a virtual water consumption rate sample is subjected to data enhancement; priori knowledge is supplemented, and a water level prediction deep learning model of an LSTM-assumed generalized physical model is established; total error control simulation and similar process simulation are added, a virtual water consumption rate bias coefficient is calibrated, and a model is applied to predict the water level; the water level prediction accuracy and reliability of the runoff hydropower station are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrological prediction, in particular to a method for predicting the water level of a runoff hydropower station with hypothetical data augmentation and virtual parameters. Background Technique

[0002] In water resource management and hydropower station operation, accurately predicting the water level of a runoff hydropower station is of great significance for improving energy production efficiency, reducing flood risks, and optimizing water resource allocation. However, due to the fact that the hydrological process of a runoff hydropower station is affected by various complex factors, including rainfall, upstream inflow, and unit operation and output arrangement, etc., traditional hydrological models face huge challenges in predicting the water level. In recent years, with the development of deep learning technology, data-driven methods have shown good application prospects in hydrological prediction. Deep learning models have the ability to automatically extract data features and capture complex non-linear relationships, and can, to a certain extent, make up for the deficiencies of traditional physical models. However, simply relying on data-driven methods may ignore the physical laws in the hydrological process, resulting in weak generalization ability of the model under unknown conditions. Summary of the Invention

[0003] The purpose of the present invention is to overcome the above deficiencies and provide a method for predicting the water level of a runoff hydropower station with hypothetical data augmentation and virtual parameters, which combines deep learning with a physical generalization model, that is, a physics-guided deep learning method. It can, on the basis of retaining the prior knowledge of the physical model, make full use of the advantages of deep learning to improve the accuracy and reliability of predicting the water level of a runoff hydropower station.

[0004] To solve the above technical problems, the technical solution adopted by the present invention is: a method for predicting the water level of a runoff hydropower station with hypothetical data augmentation and virtual parameters, including the following steps:

[0005] Step 1: Screen recent measured data and data of similar dispatching processes, and set boundaries;

[0006] Step 2: Establish an overall inflow proportion relationship that ignores the intraday distribution, and solve the generalized intraday inflow;

[0007] Step 3: Establish a physical generalization model based on a virtual water consumption rate, and propose a virtual water consumption rate limitation rule;

[0008] Step 4: Propose multiple assumptions, and augment virtual water consumption rate samples according to the assumption rules and the virtual water consumption rate model;

[0009] Step 5: Supplement prior knowledge and establish a deep learning model for predicting the water level of an LSTM - assumed generalization physical model;

[0010] Step 6: Add the simulation of total error control and the simulation of similar processes, calibrate the bias coefficient of the virtual water consumption rate, and use the model to predict the water level.

[0011] Further, the specific process of the above Step 1 is as follows:

[0012] Step 1.1: Extract recent hydrological data: Extract the upstream flow, the upstream water level of the runoff hydropower station, and the outflow flow within two weeks; unify the time scale to the minute level. For data sources at the hourly level, if the upstream boundary condition is a natural flow process, it is affected by natural rainfall runoff, resulting in a large degree of fluctuation in the flow data process. To ensure data smoothing, the corresponding hydrological data is interpolated using the spline interpolation method; for each interval [t i , t i+1 , the cubic spline interpolation function is:

[0013] f(t) = a i + b i (t - t i ) + c i (t - t i ) 2 + d i (t - t i ) 3 (1);

[0014] In the formula, t is the time of the interpolation point, t i is the start time of the interval, and a, b, c, and d are the cubic spline coefficients;

[0015] For the cascade reservoir data with the upstream boundary condition being a reservoir, which is affected by reservoir regulation and limited by the daily water level variation requirement specified in the regulations, the data fluctuation is small and it is more suitable for the linear interpolation method. The linear interpolation formula is:

[0016]

[0017] In the formula, f(t i ) and f(t i+1 ) are the flow rates or water levels of adjacent known data points;

[0018] Step 1.2: Extract similar hydrological processes: Using the day as the time scale for similar queries and 15 minutes as the time scale for data discretization, the Pearson correlation coefficient method is used to extract the similar processes of the days with the largest correlation coefficient in the past year; the formula for calculating the target flow process and the historical flow process using the correlation coefficient method is:

[0019]

[0020] In the formula, x i and y iare the values of the target flow process and the historical flow process at the $i$-th moment respectively, and are the means of the target flow process and the historical flow process respectively;

[0021] For the water level similarity process, it is required that the daily average water level in front of the dam of the runoff hydropower station differs from the target control daily average water level within the prediction period by less than 0.5 meters;

[0022] Step 1.3, boundary setting: Set boundary conditions including: initial water level at the beginning of the day, upstream predicted runoff generation and confluence, and water level-storage curve.

[0023] Furthermore, the said Step 2 includes the following contents: Obtain the upstream daily average flow magnitude, calculate the corresponding propagation lag time at this magnitude, use this magnitude of propagation lag time to represent the propagation lag time at each time point of the day, extract the total flow volume in the corresponding time periods of the upstream and downstream. For the upstream reservoir, directly extract the total flow volume of the downstream reservoir's inflow corresponding to the upstream output, calculate the conversion ratio, and determine the inflow of the assumed intra-day distribution in each period according to the daily conversion ratio.

[0024] Even further, the said Step 2 specifically includes the following process:

[0025] Step 2.1, determine the upstream inflow magnitude, with the daily average flow as the judgment basis;

[0026] In Step 2.2, when determining the water flow propagation time, if there are measured flow monitoring data, perform classification processing based on the data; appropriately merge and generalize the water flow propagation time corresponding to different flow levels to ensure that the generalized time interval is less than 10 minutes; if there are no measured data, use a mathematical model to estimate the propagation duration; when selecting the estimation method, fully consider the time cost and the river channel terrain type; for areas with a relatively short river channel length, avoid using the full solution of the complex Saint-Venant equation, but use a simplified model to reduce the calculation time and resource consumption; in the case of no data support, use a model with terrain adaptability and cost-effectiveness, and solve the estimation through the Manning formula. The formula is:

[0027]

[0028] where $v$ is the flow velocity, $n$ is the Manning roughness coefficient, $R$ is the hydraulic radius, and $S$ is the hydraulic gradient;

[0029] The specific solution and estimation process is:

[0030] 1) Determine the length, gradient, and cross-sectional shape of the river channel along the way; the gradient takes the average gradient, and the cross-sectional shape selects a typical cross-sectional shape to represent the cross-sections along the way; solve the hydraulic radius corresponding to different water depths quickly;

[0031] 2) Assume the initial conditions: The initial water depth is set to y. According to the typical cross-sectional shape, the cross-sectional area A and the wetted perimeter P can be obtained. Then, the calculation formula for the hydraulic radius R is:

[0032]

[0033] Substituting the Manning formula into the flow formula, we get:

[0034]

[0035] In the formula, S is the hydraulic gradient, and n is the Manning roughness coefficient. For general riverbed topography, according to its composition, vegetation, and shape, its general value is used to replace the whole-course value. For natural rivers of the riverbed:

[0036] Fine sand: n ≈ 0.020;

[0037] Medium gravel: n ≈ 0.025 to 0.030;

[0038] Coarse gravel: n ≈ 0.035 to 0.040;

[0039] 3) Iterative solution: After updating the water depth y, calculate Q and compare it with the known flow rate until Q is close to the known flow rate level. Determine y and A, and calculate the average flow velocity; use the river length and the average flow velocity of the river to calculate the water flow propagation duration;

[0040] Step 2.3: According to the propagation duration, select the flow rate data corresponding to the upstream and downstream time periods. Without considering the factor of water flow propagation attenuation, calculate the overall proportional conversion coefficient, and calculate the time-sharing downstream inflow according to the overall proportional conversion coefficient; when predicting the upstream flow rate process in the future, use the same method to calculate the daily generalized inflow.

[0041] Furthermore, the said Step 3 includes the following contents: Under the condition that there is generalization in the water level prediction model, through water volume balance calculation, use the generalized value to calculate the water consumption rate of the non-real runoff hydropower station; the virtual water consumption rate values at each time point have obvious deviations from the real water consumption rate values, and the deviation amount is the error existing in the generalization model corresponding to each time point. The error includes the deviation caused by generalization and the systematic measurement error; its limitation rule is: According to its definition, under the condition of satisfying the water volume balance in the local time period, smoothing processing can be carried out to eliminate the influence of outliers.

[0042] Even further, the said Step 3 specifically includes the following process:

[0043] Step 3.1: Apply the polynomial fitting method to fit the water level - storage capacity curve: The quadratic polynomial simulates the non - linear relationship between the water level and the storage capacity, and the fitting formula is:

[0044] V = a + bH + cH 2 (7);

[0045] Where V is the storage capacity, H is the water level, and a, b, and c are model parameters;

[0046] Data fitting is performed by the least squares method, and finally the model parameters are optimized to minimize the error between the actual storage capacity and the storage capacity predicted by the model;

[0047] Step 3.2: Apply the fitting formula to calculate the change in storage capacity corresponding to the change in water level at adjacent time intervals, and calculate the average flow change during the corresponding time interval;

[0048] Step 3.3: According to the water balance, apply the daily generalized inflow, the average flow change during the time interval, and the historical output process of the run-of-river hydropower station to calculate the virtual water consumption rate corresponding to each time point;

[0049] Q = C * P (8);

[0050] Where Q is the outflow, C is the virtual flow water consumption rate, and P is the power station output;

[0051] Step 3.4: Conditional smoothing and control of the total water volume error: Since multiple generalization and fitting steps result in the virtual water consumption rate inheriting the errors of different steps, there will be large fluctuations for similar processes, which is not conducive to subsequent learning. It is necessary to smooth the process and ensure the rationality of the smoothing; Based on the historical output process of the run-of-river hydropower station, smooth the virtual water consumption rate under similar output conditions. The following two methods are respectively selected according to the differences of similar conditions:

[0052] 1) For the process with a flat control water level and small output fluctuations in a long time period, the virtual water consumption rate of the time period is the average value of this period; After calculating the average value, calculate the outflow of the time period in reverse, and balance the total water volume within the time period;

[0053] 2) For processes that are similar in a short time or have a large water level variation range, use the moving average method to remove the short-term fluctuations in the time series data and highlight its long-term trend; The moving smoothing method averages the data points according to the window size n. Here, to avoid the excessive adverse effects of the fluctuations at the far points on the smoothing here, the weighted moving average method is used, and different weights are assigned according to the distances of the data points in the window. The formula is:

[0054]

[0055] Where p i is the data point in the i-th period, and w i is the weight in the i-th period.

[0056] Furthermore, the specific process of step 4 includes the following:

[0057] Assumption 1: There are systematic errors in each calculation step that can be studied; Assumption 2: Under the condition of similar upstream reservoir output or outflow process, the virtual water consumption rate is approximately constant; Assumption 3: When the water level in front of the runoff hydropower station fluctuates, the virtual water consumption rate is approximately negatively correlated with it; The implementation steps are as follows:

[0058] Step 4.1: Extract the source data, perform data cleaning, handle missing values and outliers, and perform rule verification on the data; Obtain the average propagation duration within the required time period, keep the propagation time interval unchanged, and establish a water balance equation using the water level-storage curve and information such as inflow and outflow and reservoir water level.

[0059] Step 4.2: Batch synchronously modify the upstream flow rate. According to the water balance relationship in the corresponding time period, calculate the increased or decreased water volume in the system. Obtain the flow water consumption rate at the corresponding time according to Assumption 2, and modify the actual output process to increase or decrease the outflow to make up for the water volume error.

[0060] Step 4.3: Extract the modified flow rate process of the time period as the data result after data enhancement; According to the analysis of the fluctuation law of the water consumption rate, determine that the value of modifying the upstream flow rate does not exceed 5‰ of the original data, and various enhancement methods can be introduced, including random fluctuation, batch fluctuation, and symmetric fluctuation of the front and back time periods.

[0061] Step 4.4: Make a minor amplitude modification to the water level. The virtual water consumption rate is slightly corrected according to the head difference for the fluctuated water level, and calculate the power generation process of the power station according to the water balance equation.

[0062] Furthermore, the specific process of the said Step 5 is as follows:

[0063] The input variables and model formula of the water level prediction deep learning model of the LSTM-Assumed Generalized Physical Model are:

[0064] Input variables:

[0065] Upstream flow rate process, including flow rate data from time points t - N + 1 to t;

[0066] Power generation process of the hydropower station, including output data from time points t - N + 1 to t;

[0067] Reservoir water level at the current moment;

[0068] Virtual power generation flow water consumption rate at the current moment;

[0069] The synthesis formula of the fusion model is:

[0070] Forget gate:

[0071]

[0072] Input gate:

[0073]

[0074] Cell state update:

[0075]

[0076] Output gate:

[0077]

[0078] h t = o t *tanh(C t ) (15);

[0079] The loss function formula of the prior knowledge is as follows:

[0080]

[0081] In the formula, f is the forget gate, σ represents the Sigmoid function, i is the input gate, o is the output gate, the candidate memory state, C is the memory state, h is the hidden state, W f 、W i 、W c 、W o are the weight matrices corresponding to the forget gate, input gate, candidate state generation, and output gate respectively, b f 、b i 、b c 、b o are the bias vectors corresponding to the forget gate, input gate, candidate state generation, and output gate respectively, t, t - 1, t - N + 1 are different corresponding times respectively, L original is the original model loss, λ is the adjustment balance coefficient for adjusting the relative importance of prior knowledge and the original loss, and max is the maximum value function;

[0082] When the virtual power generation flow water consumption rate in the formula is within the range defined by the prior knowledge, the loss function degenerates into the basic loss function. If it exceeds the upper limit or is lower than the lower limit, the loss function adds a constraint regularization term to punish the paranoid behavior;

[0083] The implementation steps of the LSTM - assumed generalized physical model water level prediction deep learning model are as follows:

[0084] Step 5.1. Data Processing: Extract features from the original data and the data after data augmentation using the sliding window technique to create historical window data for upstream flow and hydropower station output; perform standardization processing on the data to improve the convergence speed and performance of the model during training;

[0085] Step 5.2. Define the LSTM network structure, determine the number of layers of the LSTM network and the number of neurons in each layer, and randomly initialize the weight and bias parameters of the network; write the loss function, including the original loss term and the constraint regularization term;

[0086] Step 5.3. Dynamically adjust the learning rate and adopt the learning rate decay strategy during training, introduce L2 regularization and Dropout layer to reduce the risk of model overfitting; use the grid search method to refine the hyperparameters of the LSTM network, including the number of layers and the number of neurons, to ensure that the model can achieve the required accuracy and maintain simplicity as much as possible; in addition, implement the early stopping strategy to avoid overtraining the model on the augmented dataset, ensure good generalization ability, and prevent overfitting caused by overly complex structure or excessive learning;

[0087] Step 5.4. Calculate the change in reservoir water level using the virtual water consumption rate learned by the model; predict and monitor the dynamic change of the reservoir water level through the water balance equation to ensure the accuracy of the model output; the learning effect of the model will be evaluated by comparing the predicted reservoir water level with the actual observed value, so as to verify the prediction ability and practical application value of the model.

[0088] Further, the implementation process of adding the total error control in Step 6 is as follows: Through the numerical statistical law introduced by prior knowledge, after the model prediction calculation, further apply this law to correct the overall prediction result. The correction method is to add a bias correction coefficient to the predicted virtual water consumption rate, and do not adjust the virtual water consumption rate value of a certain part alone, but increase a fixed bias amount as a whole; the similar process simulation process in Step 6 is as follows: Select a historical process similar to the next-day prediction process, apply the fusion model for prediction, compare the prediction result with the actual process, and judge the comparison result based on multiple statistical values such as the end-of-day result and the mean value, and further refine and correct the bias coefficient.

[0089] Advantages of the present invention:

[0090] 1. The method of the present invention can effectively reduce the cumulative error of runoff hydropower station water level prediction; this method is especially suitable for the conditions where the operation data of runoff hydropower stations is less or the data reliability is limited, and can quickly and accurately predict the relationship between power generation and water level;

[0091] 2. The present invention improves the problem of prediction cumulative error caused by the sensitive head effect in the traditional physical generalization model for a runoff hydropower station, and at the same time solves the limitation of the prediction ability of previous deep learning models such as LSTM under the conditions of small data volume or limited data reliability;

[0092] 3. The present invention introduces the concept of virtual flow water consumption rate. On the basis of retaining the physical generalization model, this concept incorporates multi-step systematic errors into the virtual water consumption rate parameter, so as to better utilize the advantages of the deep learning model and avoid the problems of multi-step deep learning amplifying step-by-step errors or only relying on direct learning prediction results while ignoring data characteristics;

[0093] 4. The present invention takes the physical model as the basis for overall calculation and prediction, and restricts the prediction results of deep learning. Generally, the prediction result is the result of deep coupling between the physical model and the deep learning model, and in the worst case, the model degrades to a pure physical model. This strategy significantly improves the reliability and prediction efficiency of the model.

[0094] 5. The present invention combines deep learning with the physical generalization model, that is, adopts a physics-guided deep learning method, which can make full use of the advantages of deep learning on the basis of retaining the prior knowledge of the physical model, and improve the accuracy and reliability of water level prediction for runoff hydropower stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Attached Figure 1 is a framework diagram of a water level prediction method for a runoff hydropower station with hypothetical data augmentation and virtual parameters;

[0096] Attached Figure 2 is a schematic diagram of the result of the LSTM method learning the true virtual water consumption rate;

[0097] Attached Figure 3 is the daily application result of the prediction method based on hypothetical data augmentation and virtual parameters during the non-gate-opening period;

[0098] Attached Figure 4 is the prediction result of the coupling model for sudden plan changes;

[0099] Attached Figure 5 is the prediction result of the coupling model during the flood season. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0100] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0101] Embodiment 1:

[0102] As Figure 1 shown, a water level prediction method for a runoff hydropower station with hypothetical data augmentation and virtual parameters includes the following steps:

[0103] Step 1: Screen recent measured data and data of similar scheduling processes, and set boundaries;

[0104] Step 2: Establish the overall proportion relationship of the inflow rate into the reservoir that ignores the intraday distribution, and solve the generalized intraday inflow rate into the reservoir;

[0105] Step 3: Establish a physical generalization model based on the virtual water consumption rate, and propose rules for limiting the virtual water consumption rate;

[0106] Step 4: Propose multiple assumptions, and enhance the virtual water consumption rate samples according to the assumption rules and the virtual water consumption rate model;

[0107] Step 5: Supplement prior knowledge and establish a deep learning model for water level prediction of the LSTM - assumption generalization physical model;

[0108] Step 6: Add error total control simulation and similar process simulation, calibrate the bias coefficient of the virtual water consumption rate, and apply the model to predict the water level.

[0109] In this embodiment, for the screening and boundary setting of recent measured data and data of similar scheduling processes in Step 1, the time period of the recent measured data is 2 to 3 weeks, the time resolution is 15 minutes to 1 hour, and the data types are: the outflow rate of the upstream reservoir (control site) (or the output of the upstream cascade reservoir, the water level in front of the dam), the inflow rate into the run - of - river hydropower station, the total plant output, the outflow rate, the water level in front of the dam, and the water level behind the dam. The similar scheduling process is: taking the line type and mean value of the total plant output of the run - of - river hydropower station as the judgment benchmark for the similar line type, determining the flow similar process according to the similarity coefficient, and approximately finding the daily scheduling process in which the difference between the initial water level at the start of the day in front of the dam of the run - of - river hydropower station and the intraday water level mean value in the historical data similar line type scheduling process is within 0.5 m. The boundary settings are: the initial water level at the start of the day, the daily outflow rate of the upstream control site.

[0110] Further, the specific process of Step 1 is as follows:

[0111] Step 1.1: Extract recent hydrological data: Extract the upstream flow (output), the upstream water level of the run - of - river hydropower station, and the outflow rate within two weeks; unify the time scale to the minute level. For data sources at the hourly level, if the upstream boundary condition is a natural flow process, which is affected by natural rainfall runoff and causes a large fluctuation in the flow data process, in order to ensure data smoothing, the corresponding hydrological data is interpolated using the spline interpolation method. For each interval [t i ,t i+1 , the cubic spline interpolation function is:

[0112] f(t) = a i +b i (t - t i )+c i (t - ti ) 2 +d i (t - t i ) 3 (1);

[0113] Wherein, t is the time of the interpolation point, t i is the start time of the interval, and a, b, c, d are cubic spline coefficients;

[0114] For the cascade reservoir data with the upstream boundary condition being a reservoir, affected by reservoir regulation and limited by the daily water level variation requirement of the regulations, the data fluctuates less and is more suitable for the linear interpolation method. The linear interpolation formula is:

[0115]

[0116] Wherein, f(t i ) and f(t i+1 ) are the flow rates or water levels of adjacent known data points.

[0117] Step 1.2. Extract similar hydrological processes: Using the day as the time scale for similar queries and 15 minutes as the time scale for data discretization, adopt the Pearson correlation coefficient method to extract the similar processes of the days with the largest correlation coefficient in the past year; the formula for calculating the correlation coefficient between the target flow process and the historical flow process is:

[0118]

[0119] Wherein x i and y i are respectively the values of the target flow process and the historical flow process at the i-th moment, and are respectively the means of the target flow process and the historical flow process;

[0120] For the water level similar process, it is required that the daily average water level in front of the runoff hydropower station dam differs from the target control daily average water level within the prediction period by less than 0.5 meters.

[0121] Step 1.3. Boundary setting: Set boundary conditions including: the initial water level at the beginning of the day, upstream predicted runoff generation and confluence (future reservoir output plan), and water level - storage capacity curve.

[0122] Furthermore, the said Step 2 includes the following contents: Obtain the upstream daily average flow rate magnitude, calculate the corresponding propagation lag time at this magnitude, use this magnitude propagation lag time to represent the propagation lag time of each time point on the current day, extract the total flow rates of the corresponding time periods upstream and downstream, directly extract the total flow rate of the downstream reservoir inflow corresponding to the upstream output for the upstream reservoir, calculate the conversion ratio, and determine the assumed daily - distributed inflow for each time period according to the daily conversion ratio.

[0123] Further, the specific process of Step 2 includes the following:

[0124] Step 2.1: Determine the upstream incoming water flow level, with the daily average flow as the judgment basis.

[0125] Step 2.2: When determining the water flow propagation time, if there are measured flow monitoring data, grading treatment should be carried out according to the data. The water flow propagation times corresponding to different flow levels can be appropriately combined and generalized to ensure that the generalized time interval is less than 10 minutes. If there are no measured data, a mathematical model should be used to estimate the propagation duration. When selecting the estimation method, time cost and river channel terrain type should be fully considered. For areas with a short river channel length, the complex full solution of the Saint-Venant equation can be avoided, and a simplified model can be used to reduce calculation time and resource consumption. This method can improve the estimation efficiency while maintaining reasonable accuracy. In the case of no data support, a model with terrain adaptability and cost-effectiveness can be used, such as solving the estimation using the Manning formula. The formula is:

[0126]

[0127] In the formula, v is the flow velocity, n is the Manning roughness coefficient, R is the hydraulic radius, and S is the hydraulic gradient.

[0128] The specific solution and estimation process is as follows:

[0129] 1) Determine the length, gradient, and cross-sectional shape of the river channel along the way; the gradient takes the average gradient, and the cross-sectional shape selects a typical cross-sectional shape to represent the cross-sections along the way; solve the hydraulic radius corresponding to different water depths quickly.

[0130] 2) Assume the initial conditions: Initially set the water depth as y. According to the typical cross-sectional shape, the cross-sectional area A and the wetted perimeter P can be obtained. Then the calculation formula for the hydraulic radius R is:

[0131]

[0132] Substitute the Manning formula into the flow formula to get:

[0133]

[0134] In the formula, S is the hydraulic gradient, and n is the Manning roughness coefficient. For general river channel terrain, according to its composition, vegetation, and shape, its general value is used to replace the whole-process value. For the natural river channel of the riverbed:

[0135] Fine sand: n ≈ 0.020;

[0136] Medium gravel: n ≈ 0.025 to 0.030;

[0137] Coarse gravel: n ≈ 0.035 to 0.040;

[0138] 3) Iterative solution: After updating the water depth y, calculate Q and compare it with the known flow rate. Keep calculating until Q is close to the known flow rate level. Determine y and A, and calculate the average flow velocity. Use the river channel length and the average flow velocity of the river channel to calculate the water flow propagation duration.

[0139] Step 2.3: According to the propagation duration, select the flow rate data corresponding to the upstream and downstream time periods. Without considering the factor of water flow propagation attenuation, calculate the overall proportional conversion coefficient, and calculate the time-sharing downstream inflow according to the overall proportional conversion coefficient. When predicting the upstream flow rate process in the future, use the same method to calculate the daily generalized inflow.

[0140] Furthermore, the definition of the virtual water consumption rate in step 3 is as follows: Under the condition that there are generalizations in the water level prediction model (such as the conversion of the upstream reservoir output into the outflow, and the attenuation and energy dissipation characteristics in the non-steady flow propagation characteristics), through the water balance calculation, use the non-realized water consumption rate of the runoff hydropower station calculated by the generalized numerical calculation. Its characteristics are: The virtual water consumption rate values at each time point have obvious deviations from the real water consumption rate values, and the deviation amount is the error existing in the generalized model corresponding to each time point. The error includes the deviation caused by generalization and the systematic measurement error. The characteristic of its limitation rule is: According to its definition, under the condition of satisfying the water balance in the local time period, smoothing processing can be carried out to eliminate the influence of outliers.

[0141] Even further, step 3 specifically includes the following processes:

[0142] Step 3.1: Apply the polynomial fitting method to fit the water level - storage capacity curve. The quadratic polynomial can better simulate the non - linear relationship between the water level and the storage capacity. The fitting formula is:

[0143] V = a + bH + cH 2 (7);

[0144] Where V is the storage capacity, H is the water level, and a, b, c are the model parameters;

[0145] Data fitting is carried out by the least - squares method, and finally the model parameters are optimized to minimize the error between the actual storage capacity and the storage capacity predicted by the model;

[0146] Step 3.2: Apply the fitting formula to calculate the change in storage capacity corresponding to the change in water level at adjacent time intervals, and calculate the average flow rate change within the corresponding time interval.

[0147] Step 3.3: According to the water balance, apply the daily generalized inflow, the average flow rate change within the time interval, and the historical output process of the runoff hydropower station to calculate the virtual water consumption rate corresponding to each time point;

[0148] Q = C * P (8);

[0149] Where Q is the out - flow rate, C is the virtual flow water consumption rate, and P is the power generation of the power station;

[0150] Step 3.4: Smooth the conditions and control the total water volume error. Since the multiple generalization and fitting steps cause the virtual water consumption rate to inherit the errors of different steps, there will be large fluctuations for similar processes, which is not conducive to subsequent learning. It is necessary to smooth the process and ensure the rationality of the smoothing. According to the historical power generation process of the run - of - river hydropower station, smooth the virtual water consumption rate under similar power generation conditions. The following two methods are respectively selected according to the differences of similar conditions:

[0151] 1) For the process with stable regulated water level and small power generation fluctuations in a long time period, the virtual water consumption rate of each time period adopts the average value of this period. After calculating the average value, calculate the out - flow rate of each time period backward to ensure the overall water volume balance within the time period.

[0152] 2) For the process with short - term similarity or large water level fluctuations, use the moving average method to remove the short - term fluctuations in the time - series data and highlight its long - term trend. The moving smoothing method averages the data points according to the window size n. Here, to avoid the excessive adverse effects of far - point fluctuations on the smoothing here, the weighted moving average (WMA) method is adopted. Different weights are assigned according to the distances of each data point in the window. The formula is:

[0153]

[0154] In the formula, p i is the data point of the i - th period, and w i is the weight of the i - th period. Usually, the weights are increasing, such as w i = i. For the data of n periods, the later data has a greater influence.

[0155] Furthermore, the specific process of step 4 includes the following:

[0156] Assumption 1: There are systematic errors in each calculation step that can be studied; Assumption 2: Under the conditions of similar upstream reservoir output or outflow process, the virtual water consumption rate is approximately constant; Assumption 3: When the water level in front of the runoff hydropower station fluctuates, the virtual water consumption rate is approximately negatively correlated with it; According to the assumption rules and the data enhancement of the virtual water consumption rate model, its enhancement algorithm is as follows: ① Select the historical process, artificially unify or randomly fluctuate the upstream reservoir outflow (output), and the fluctuation range does not exceed 5‰ of the current value. The virtual water consumption rate remains unchanged, and the new outflow is obtained according to the water balance. The fluctuation error will be synchronously integrated and uniformly learned and identified by the subsequent deep learning model. ② Select the historical process, artificially unify or randomly fluctuate the water level in front of the runoff hydropower station, and the fluctuation range does not exceed 1 cm. The virtual water consumption rate is adjusted according to the fluctuation value ratio, and the adjustment ratio does not exceed the ratio of the fluctuation range to the total head of the runoff hydropower station. Then, the new outflow is calculated according to the water balance.

[0157] The implementation steps are specifically as follows:

[0158] Step 4.1: Extract the source data, perform data cleaning, handle missing values and outliers, and perform rule verification on the data; Obtain the average propagation duration within the required time period, keep the propagation time interval unchanged, and establish a water balance equation using the water level-storage curve and information such as inflow and outflow, and reservoir water level.

[0159] Step 4.2: Batch-synchronously modify the upstream flow rate, calculate the increased or decreased water volume in the system according to the water balance relationship in the corresponding time period, obtain the flow water consumption rate at the corresponding time according to Assumption 2, and make up for the water volume error by modifying the actual output process and increasing or decreasing the outflow.

[0160] Step 4.3: Extract the modified flow rate process during the time period as the data result after data enhancement; Determine according to the analysis of the water consumption rate fluctuation law that the value of modifying the upstream flow rate does not exceed 5‰ of the original data, and various enhancement methods can be introduced, including random fluctuation, batch fluctuation, and symmetric fluctuation of the front and back time periods.

[0161] Step 4.4: Make a slight change in the water level, slightly correct the virtual water consumption rate according to the head difference for the floating water level, and calculate the power generation process of the power station according to the water balance equation.

[0162] The prior knowledge limitation content in step 5 is as follows: the statistical characteristics of historical data, such as the hydropower station output corresponding to the daily average flow, and artificial experience. The establishment process of the LSTM-hypothetical generalized physical model is as follows: using the long short-term memory network (LSTM) as the core algorithm, by analyzing the time series characteristics of the boundary input data, the predicted water level state is obtained. Then, through the assumed physical generalization model, the directly predicted water level state is converted into an intermediate step virtual value - virtual water consumption rate calculated by the physical generalization model. The virtual water consumption rate time series and the boundary conditions (upstream and downstream flow (output), upstream and downstream water levels) of the corresponding time series are used to construct learning training and test data through the sliding window method, adjust the model structure and verify, and a stable fusion model is formed after multiple iterative tests.

[0163] Furthermore, the specific process of step 5 is as follows:

[0164] The input variables and model formula of the LSTM-hypothetical generalized physical model's water level prediction deep learning model are:

[0165] Input variables:

[0166] Upstream flow process, including flow data from time point t - N + 1 to t;

[0167] Hydropower station output process, including output data from time point t - N + 1 to t;

[0168] The reservoir water level at the current moment;

[0169] The virtual power generation flow water consumption rate at the current moment;

[0170] The synthesis formula of the fusion model is:

[0171] Forget gate:

[0172]

[0173] Input gate:

[0174]

[0175] Cell state update:

[0176]

[0177] Output gate:

[0178]

[0179] h t =o t*tanh(C t ) (15);

[0180] The loss function formula of the prior knowledge is as follows:

[0181]

[0182] In the formula, f is the forget gate, σ represents the Sigmoid function, i is the input gate, o is the output gate, the candidate memory state, C is the memory state, h is the hidden state, W f 、W i 、W c 、W o are the weight matrices corresponding to the forget gate, input gate, candidate state generation, and output gate respectively, b f 、b i 、b c 、b o are the bias vectors corresponding to the forget gate, input gate, candidate state generation, and output gate respectively, t, t - 1, t - N + 1 are different corresponding moments respectively, L original is the original model loss, λ is the adjustment balance coefficient for adjusting the relative importance of prior knowledge and the original loss, and max is the maximum value function;

[0183] When the virtual power generation flow water consumption rate in the formula is within the range defined by the prior knowledge, the loss function degenerates into the basic loss function. If it exceeds the upper limit or is lower than the lower limit, the loss function adds a constraint regularization term to punish the paranoid behavior;

[0184] The implementation steps of the LSTM - assumed generalized physical model water level prediction deep learning model are as follows:

[0185] Step 5.1, Data processing: Extract features from the original data and the data after data augmentation using the sliding window technique to create historical window data for upstream flow and hydropower station output; Standardize the data to improve the convergence speed and performance of the model during training;

[0186] Step 5.2, Define the LSTM network structure, determine the number of layers of the LSTM network and the number of neurons in each layer, and randomly initialize the weight and bias parameters of the network; Write the loss function, including the original loss term and the constraint regularization term;

[0187] Step 5.3: Dynamically adjust the learning rate during training and adopt a learning rate decay strategy. Introduce L2 regularization and Dropout layers to reduce the risk of model overfitting. Use the grid search method to refine the hyperparameters of the LSTM network, including the number of layers and neurons, to ensure that the model can achieve the required accuracy while maintaining simplicity as much as possible. In addition, implement an early stopping strategy to avoid overtraining the model on the enhanced dataset, ensure good generalization ability, and prevent overfitting caused by overly complex structures or excessive learning.

[0188] Step 5.4: Use the virtual water consumption rate learned by the model to calculate the change in reservoir water level. Through the water balance equation, predict and monitor the dynamic change of the reservoir water level to ensure the accuracy of the model output. The learning effect of the model will be evaluated by comparing the predicted reservoir water level with the actual observed value, so as to verify the prediction ability and practical application value of the model.

[0189] Furthermore, the implementation process of adding the total error control in Step 6 is as follows: Through the numerical statistical law introduced by prior knowledge, after the model prediction calculation, further apply this law to correct the overall prediction result. The correction method is to add a bias correction coefficient to the predicted virtual water consumption rate, without adjusting the virtual water consumption rate value of a certain part alone, and increase a fixed bias as a whole. The simulation process of the similar process in Step 6 is as follows: Select a historical process similar to the next-day prediction process, apply the fusion model for prediction, compare the prediction result with the actual process, and determine the comparison result based on multiple statistical values such as the end-of-day result and the mean value, and further finely correct the bias coefficient.

[0190] The specific implementation steps are as follows:

[0191] 1. Calibration of the similar process. First, preferentially select the flow and output data of the most recent two weeks as the benchmark, and then evaluate the similarity of the historical data by applying the Pearson correlation coefficient and the dynamic time warping (DTW) method. For the historical natural measured time series data, select the Pearson correlation coefficient method to evaluate the linear correlation degree between the flow data.

[0192]

[0193] In the formula, x i and y i are the observed values of two time series respectively, and are the average values of their respective sequences

[0194] For the artificially generated data, select the DTW method to evaluate the similarity between the output process curves of the runoff power stations. Determine the similarity between sequences by calculating the shortest path distance between different time points.

[0195]

[0196] where A and B are two time series, a i and b j are elements in the series, d(a i , b j ) is the distance between elements (usually the Euclidean distance), and min means finding the path with the minimum total distance among all possible paths.

[0197] On this basis, a similarity acceptance threshold is set, and recent data that meet the requirements of higher flow similarity and output curve similarity than this threshold are preferentially screened. If the recent data is not sufficient to constitute an effective prediction, the process similar to the current situation in historical data is further searched.

[0198] 2. Total deviation control. The similar process is used as a reference for the learning process, and the total deviation control is to limit the overall error. Especially in the scenario of continuous prediction for a run-of-river hydropower station, the originally small error will be quickly amplified. By correcting the initial value and combining the total deviation control method, the usability of the model is ensured. The total deviation control method is relatively simple. According to the assumed daily inflow and outflow magnitude, the historical daily output range corresponding to this flow magnitude is evaluated from historical data. After obtaining a rough reference, the predicted water level corresponding to the daily output is corrected by adjusting the overall prediction size of the virtual flow water consumption rate.

[0199] Example 2:

[0200] Apply the coupling model and the solution method to a real production work case. Taking the Three Gorges - Gezhouba as the research area, the two cascade reservoirs are 38 kilometers apart and there are no large tributaries in the middle. The outflow of the Three Gorges Reservoir is the inflow of the Gezhouba Reservoir. The Gezhouba Water Control Project is designed as the shipping cascade and anti-regulation reservoir of the Three Gorges Water Control Project. As a run-of-river hydropower station, it only has daily regulation capacity itself, and the flow data between the Three Gorges and Gezhouba are calculated using the unit curve data of their respective units, resulting in differences in flow data and the "water volume imbalance" phenomenon between the Three Gorges and Gezhouba. These conditions respectively meet the calculation requirements of the present invention, that is, the run-of-river hydropower station is too sensitive to the generating head, the dispatching plan affects the head change, gradually generating cumulative errors, and the data reliability is limited, so it is not suitable to directly use the deep learning algorithm for direct fitting simulation.

[0201] Since the production cycle of the daily regulation reservoir plan is a single day, and the test time of the invention application case is on a daily scale, according to the 96-point demand of power generation dispatching, the time resolution of the model is set to 15 minutes. There are sufficient observation data and relatively mature research results on the propagation duration between the Three Gorges and Gezhouba. The larger the outflow of the Three Gorges, the less the propagation time. When the outflow is above 7000 m 3 / s, the propagation time is basically 20 - 30 min; when the outflow is 7000 m3 Below / s, the propagation time is basically 30 - 40 min. In daily dispatching calculation and analysis, the propagation time can be calculated as 30 min.

[0202] First, apply LSTM to test the learning error of the actual water consumption rate. The test results are as Figure 2 shown. The training set R 2 is 0.99, and the test set R 2 is 0.97. The average error of a single prediction is only 0.3. However, after being tested by the physical model, the error gradually accumulates. With the water level difference, the error of the water consumption rate will increase rapidly, and the model results gradually deviate from the real process.

[0203] Further, use the method of this invention for prediction. First, ignore the errors in the attenuation process, conversion coefficient, and propagation time generalization. Calculate the virtual water consumption rate. Take January 24th, January 26th, and May 4th, 2024 as typical representatives, which mainly represent the sudden dispatching process of temporary peak regulation, the dispatching process in the general dry season, and the dispatching process in the general flood season. When making the next-day dispatching plan, each case selects the measured data of the previous three weeks as the research basis. After data augmentation based on assumptions, some data are extracted as the test set, and the rest are the training set. When the training set R 2 reaches above 0.99 and the test set R 2 reaches above 0.9, the early stopping mode is enabled. By making the 96-point plan for the next day, predict the water level control process for the next day. The comparison between the measured output process and the planned output process, and the comparison between the predicted water level change process and the measured water level process are respectively as Figures 3 to 5 shown. The results show that during the non-flood season, the daily average water level prediction error is about 10 cm. During the flood season, the error increases, but the error range can still maintain the accuracy of plan making.

[0204] The following conclusions are obtained from this comparison case:

[0205] Conclusion 1: In the case of limited data reliability, the mode of directly predicting the water level by the deep learning method is restricted. However, by calculating the parameters step by step, better results can still be obtained using deep learning, but there are relatively large cumulative errors in the overall calculated water level prediction results.

[0206] Conclusion 2: The method effectively controls the cumulative error and reduces the influence of limited data reliability on the prediction results.

[0207] Conclusion 3: During the flood season, due to insufficient sample quantity and limited data reliability, the model relies more on the total deviation control and prior knowledge module in this invention.

[0208] It can be seen that the proposed method for predicting the water level of a runoff hydropower station with hypothetical data augmentation and virtual parameters can quickly and accurately predict the water level change process. The proposed concept of virtual flow water consumption rate effectively solves the problem of cumulative step-by-step prediction errors by aggregating multi-step errors into this parameter. The hypothetical data augmentation provides data support for the short-term learning and training of the model, realizes the coupling of the virtual flow water consumption rate and the deep learning method, and finally effectively controls the total error and the daily average error, making them within an acceptable range in daily applications. The present invention provides important support for the daily scheduling of runoff hydropower stations.

[0209] The above embodiments are only the preferred technical solutions of the present invention and should not be regarded as limitations on the present invention. The protection scope of the present invention should be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A water level prediction method for a run-of-river hydropower station assuming data enhancement and virtual parameters, characterized in that: The following steps are involved: Step 1: Filter recent measured data and similar scheduling process data and set boundaries; Step 2: Establish the overall inflow ratio relationship ignoring the intra-day distribution, and solve the intra-day generalized inflow; Step 3: Establish a physical generalization model based on virtual water consumption rate and propose virtual water consumption rate limitation rules; Step 4: Propose multiple assumptions, and according to the assumption rules and the virtual water consumption rate model, data enhance the virtual water consumption rate samples; Step 5: Supplement prior knowledge and establish a water level prediction deep learning model based on LSTM-assumed generalized physical model; Step 6: Add total error control simulation and similar process simulation, calibrate the virtual water consumption rate bias coefficient, and use the model to predict the water level.

2. The method for predicting water level of a run-of-river hydropower station based on assumed data enhancement and virtual parameters according to claim 1, characterized in that: The specific process of step 1 is as follows: Step 1.1, extract recent hydrological data: extract upstream flow within two weeks, upstream water level of run-of-river hydropower station, and outflow flow; unify the time scale to minute level. For hourly data sources, if the upstream boundary condition is a natural flow process, it is affected by natural rainfall, resulting in a large degree of fluctuation in the flow data process. In order to ensure data smoothness, the corresponding hydrological data are interpolated using the spline interpolation method; for each interval [t i ,t i+1 ], the cubic spline interpolation function is: f(t)=a i +b i (t-t i )+c i (t-t i ) 2 +d i (t-t i ) 3 (1); Where t is the time of the interpolation point, ti is the start time of the interval, and a, b, c, d are cubic spline coefficients; For the cascade reservoir data with the upstream boundary condition of the reservoir, it is affected by the reservoir regulation and is limited by the daily water level fluctuation requirements required by the regulations. The data fluctuation is small and is more suitable for the linear interpolation method. The linear interpolation formula is: In the formula, f(t i ) and f(t i+1 ) is the adjacent known data point flow or water level; Step 1.2, extract similar hydrological processes: take the day as the time scale of similarity query, take 15 minutes as the time scale of data discretization, use the Pearson correlation coefficient method to extract the similar processes of the few days with the largest correlation coefficient in the past year; The formula for calculating the target flow process and the historical flow process using the correlation coefficient method is: Where x i and i are the values ​​of the target flow process and the historical flow process at the i-th moment, and are the means of the target flow process and the historical flow process respectively; The water level similarity process requires that the difference between the average daily water level in front of the run-of-river hydropower station and the target control average daily water level within the prediction period should be less than 0.5 meters; Step 1.3, boundary setting: Setting boundary conditions includes: water level adjustment at the beginning of the day, upstream predicted runoff, and water level and reservoir capacity curve.

3. The method for predicting water level of a run-of-river hydropower station based on assumed data enhancement and virtual parameters according to claim 1, characterized in that: The step 2 includes the following contents: obtaining the average daily flow magnitude of the upstream, calculating the corresponding propagation lag under this magnitude, using the propagation lag of this magnitude to represent the propagation lag at each time point of the day, extracting the total flow of the corresponding time periods of the upstream and downstream, and the upstream reservoir directly extracting the total inflow flow of the downstream reservoir corresponding to the upstream output, calculating the conversion ratio and determining the assumed daily distributed inflow flow of each time period according to the daily conversion ratio.

4. The method for predicting water level of a run-of-river hydropower station based on data enhancement and virtual parameters according to claim 3 is characterized in that: The step 2 specifically includes the following process: Step 2.1, determine the upstream water flow level, using the average daily flow as the basis for judgment; Step 2.2: When determining the water flow propagation time, if there is measured flow monitoring data, perform graded processing based on the data; appropriately merge and generalize the water flow propagation time corresponding to different flow levels to ensure that the generalized time interval is less than 10 minutes; if there is no measured data, use a mathematical model to estimate the propagation time; when selecting the estimation method, fully consider the time cost and river terrain category; for areas with short river lengths, avoid using the complex full solution of the Saint-Venant equation, but use a simplified model to reduce computing time and resource consumption; in the absence of data support, use a model with terrain adaptability and cost-effectiveness to solve the estimation through the Manning formula, the formula is: In the formula, v is the flow velocity, n is the Manning roughness, R is the hydraulic radius, and S is the hydraulic slope; The specific solution and estimation process is: 1) Determine the length, slope and cross-sectional shape of the river channel; take the average slope as the slope, and select the typical cross-sectional shape to represent the cross-sectional shape along the river channel; solve the hydraulic radius corresponding to different water depths quickly; 2) Assume initial conditions: The initial water depth is set to y. According to the typical cross-sectional shape, the cross-sectional area A and the wetted perimeter P can be obtained. The calculation formula for the hydraulic radius R is: Substituting the Manning formula into the flow formula yields: Where S is the hydraulic slope and n is the Manning roughness. For general river terrain, according to its composition, vegetation, and shape, its general value should be used instead of the full value. For natural river channels: Fine sand: n≈0.020; Medium grit: n≈0.025 to 0.030; Coarse gravel: n≈0.035 to 0.040; 3) Iterative solution: After updating the water depth y, calculate Q and compare it with the known flow until Q is close to the known flow level, determine y and A, and calculate the average flow velocity; use the river length and the average flow velocity of the river to calculate the water flow propagation time; Step 2.3: According to the propagation time, select the flow data of the corresponding upstream and downstream time periods, do not consider the flattening factor of water flow propagation, calculate the overall proportional conversion coefficient, and calculate the downstream inflow flow according to the overall proportional conversion coefficient; when the upstream flow process is predicted in the future, the same method is used to calculate the daily generalized inflow flow.

5. The method for predicting water level of a run-of-river hydropower station based on assumed data enhancement and virtual parameters according to claim 1, characterized in that: The step 3 includes the following contents: when the water level prediction model is generalized, the water consumption rate of the non-real run-of-river hydropower station is calculated by using the generalized numerical value through water balance calculation; the virtual water consumption rate value at each time point has a significant deviation from the real water consumption rate value, and the deviation amount is the error existing in the generalized model corresponding to each time point, and the error includes the deviation caused by generalization and the systematic measurement error; its limiting rule is: according to its definition, when the water balance of the local time period is satisfied, smoothing processing can be performed to eliminate the influence of outliers.

6. A method for predicting water level of a run-of-river hydropower station assuming data enhancement and virtual parameters according to claim 5, characterized in that: The step 3 specifically includes the following process: Step 3.1: Use the polynomial fitting method to fit the water level and reservoir capacity curve: The quadratic polynomial simulates the nonlinear relationship between water level and reservoir capacity. The fitting formula is: V=a+bH+cH 2 (7); Where V is the reservoir capacity, H is the water level, and a, b, and c are model parameters; The data were fitted by the least square method, and finally the model parameters were optimized to minimize the error between the actual storage capacity and the storage capacity predicted by the model; Step 3.2, use the fitting formula to calculate the change in reservoir capacity corresponding to the change in water level at adjacent time intervals, and calculate the average flow change during the corresponding time interval; Step 3.3, according to the water balance, the virtual water consumption rate corresponding to each time point is calculated by applying the daily generalized inflow flow, the average flow change within the time period, and the historical output process of the run-of-river hydropower station; Q = C * P (8); Where Q is the outflow flow, C is the virtual flow water consumption rate, and P is the power station output; Step 3.4, conditional smoothing and control of total water volume error: Due to multiple generalization and fitting steps, the virtual water consumption rate will inherit the errors of different steps, which will cause large fluctuations for similar processes, which is not conducive to subsequent learning. It is necessary to smooth the process and ensure the rationality of the smoothing; according to the historical output process of the run-of-river hydropower station, the virtual water consumption rate under similar output conditions is smoothed. The smoothing method selects the following two methods according to the difference in similar conditions: 1) For processes with long-term water level control and small output fluctuations, the virtual water consumption rate for the period adopts the average value of the period; after calculating the average value, the outflow flow of the period is calculated backwards to ensure the overall balance of water volume within the period; 2) For processes with short-term similarity or large water level fluctuations, the sliding average method is used to remove short-term fluctuations in the time series data and highlight its long-term trend; the sliding smoothing method averages the data points according to the window size n. Here, in order to avoid excessive adverse effects of far-point fluctuations on the smoothing here, the weighted sliding average method is used to assign different weights according to the distance of each data point in the window. The formula is: In the formula, p i is the data point of period i, w i is the weight of the i-th period.

7. The method for predicting water level of a run-of-river hydropower station based on data enhancement and virtual parameters according to claim 1, characterized in that: The step 4 specifically includes the following process: Assumption 1: There are systematic errors in each calculation step that can be studied; Assumption 2: Under similar upstream reservoir output or outflow process conditions, the virtual water consumption rate is approximately unchanged; Assumption 3: When the water level in front of the run-of-river hydropower station fluctuates, the virtual water consumption rate is approximately negatively correlated with it; The implementation steps are: Step 4.1: Extract source data, clean the data, process missing values ​​and outliers, and perform rule verification on the data; Obtain the average propagation time in the required period, maintain the propagation time interval unchanged, and use the water level storage capacity curve and the inflow and outflow flow, reservoir water level and other information to establish the water balance equation; Step 4.2, modify the upstream flow synchronously in batches, calculate the increase or decrease of water volume in the system according to the water balance relationship in the corresponding time period, obtain the flow water consumption rate at the corresponding time according to assumption 2, and increase or decrease the outflow flow to compensate for the water volume error by modifying the actual output process; Step 4.3, extract the modified time period flow process as the data result after data enhancement; according to the data analysis of the fluctuation law of water consumption rate, determine that the value of the modified upstream flow does not exceed 5‰ of the original data, and can introduce a variety of enhancement methods, including random fluctuation, batch fluctuation, and symmetrical fluctuation of the previous and next time periods; Step 4.4: modify the water level slightly, and correct the virtual water consumption rate slightly according to the head difference after floating, and calculate the power station output process according to the water balance equation.

8. The method for predicting water level of a run-of-river hydropower station based on assumed data enhancement and virtual parameters according to claim 1, characterized in that: The specific process of step 5 is as follows: The input variables and model formula of the LSTM-water level prediction deep learning model assuming a generalized physical model are: Input variables: The upstream flow process includes flow data from time t-N+1 to time t; The power output range of the hydropower station, including the output data from t-N+1 to time point t; The reservoir water level at the current moment; The water consumption rate of virtual power generation flow at the current moment; The synthesis formula of the fusion model is: Forget Gate: Input Gate: Unit status update: Output Gate: h t =o t *tanh(C t ) (15); The loss function formula of the prior knowledge is: In the formula, f is the forget gate, σ represents the Sigmoid function, i is the input gate, and o is the output gate. Candidate memory state, C is the memory state, h is the hidden state, W f , W i , W c , W o are the weight matrices corresponding to the forget gate, input gate, candidate state generation, and output gate, respectively, f , b i , b c , b o are the bias vectors corresponding to the forget gate, input gate, candidate state generation, and output gate, respectively. t, t-1, and t-N+1 are the corresponding different moments. L original is the original model loss, λ is the adjustment balance coefficient for adjusting the relative importance of prior knowledge and original loss, and max is the maximum value function; When the water consumption rate of the virtual power generation flow in the formula is within the range limited by prior knowledge, the loss function degenerates into a basic loss function. If it exceeds the upper limit or is lower than the lower limit, the loss function adds a constrained regularization term to punish biased behavior. The implementation steps of the LSTM-water level prediction deep learning model assuming a generalized physical model are: Step 5.1, data processing: extract features from the original data and data-enhanced data using sliding window technology to create historical window data for upstream flow and hydropower station output; standardize the data to improve the convergence speed and performance of the model during training; Step 5.2: Define the LSTM network structure, determine the number of layers and neurons in each layer of the LSTM network, and randomly initialize the weights and bias parameters of the network; write the loss function, including the original loss term and the constrained regularization term; Step 5.3: Dynamically adjust the learning rate and adopt the learning rate decay strategy during the training process, introduce L2 regularization and Dropout layer to reduce the risk of model overfitting; use the grid search method to refine the hyperparameters of the LSTM network, including the number of layers and neurons, to ensure that the model can achieve the required accuracy while maintaining the greatest possible simplicity; in addition, implement the early stopping strategy to avoid overtraining of the model on the enhanced data set, ensure good generalization ability, and prevent overfitting due to overly complex structure or excessive learning; Step 5.4: Use the virtual water consumption rate learned from the model to calculate the change in reservoir water level; predict and monitor the dynamic changes of reservoir water level through the water balance equation to ensure the accuracy of the model output; the learning effect of the model will be evaluated by comparing the predicted reservoir water level with the actual observed value, thereby verifying the model's predictive ability and practical application value.

9. The method for predicting water level of a run-of-river hydropower station based on assumed data enhancement and virtual parameters according to claim 1, characterized in that: The implementation process of adding the total error control in step 6 is: the numerical statistical law introduced through prior knowledge is further applied to correct the overall prediction result after the model prediction calculation. The correction method is to add a bias correction coefficient to the predicted virtual water consumption rate, and the value of the virtual water consumption rate of a certain part is not adjusted separately, but the overall fixed bias is increased; the similar process simulation process in step 6 is: select a historical process similar to the prediction process of the next day, apply the fusion model to predict, compare the prediction result with the actual process, and use the end-of-day results and the average multiple statistical values ​​to determine the comparison result, and further finely correct the bias coefficient.

Images