Cascade pump station pre-station water level real-time prediction method based on optimized input and rolling correction

Abstract

The invention provides a cascade pump station pre-station water level real-time prediction method based on optimized input and rolling correction, and relates to the field of pump station operation scheduling and water conservancy informatization research. According to the cascade pump station front-station water level real-time prediction method based on optimization input and rolling correction, a method including collected data cleaning, input factor simplification, time-phased prediction and rolling error elimination is provided, firstly, collected related data are cleaned, and then normalization preprocessing is adopted; calculating the correlation of input factors by using a Pearson's correlation coefficient, carrying out sensitivity analysis, calculating the lag time of the water flow propagation process, carrying out time-phased prediction according to the lag time, introducing an error correction (EC) technology, inputting training set data into an LSTM-EC neural network model, and continuously adjusting and optimizing a weight matrix and a deviation item, so as to obtain a prediction result; and completing training optimization of the LSTM-EC neural network model, inputting prediction set data, and finally realizing real-time rolling prediction of the pre-station water level of the cascade pump station.

Description

Technical Field

[0001] The invention relates to the technical field of pump station operation scheduling and water conservancy information research, and specifically to a real-time prediction method for water levels in front of cascade pump stations with optimized input and rolling correction. Background Art

[0002] With the increase of population, economic development and environmental pollution, the construction of inter-basin water diversion projects has become the core of solving water shortage problems and optimizing water resource allocation. The cascade pump station system is a complex system composed of facilities such as oil pump stations, trash racks and channel sections. Accurate water level prediction is related to the safe operation of water diversion projects and provides scientific decision-making basis for dispatching personnel to control pump stations and dispatch water.

[0003] The traditional method of predicting the water level in front of cascade pumping stations is usually based on a one- and two-dimensional hydrodynamic model. Its advantage is that the model parameters have certain practical physical meanings, but the accuracy of the model requires a large amount of reasonably distributed data and appropriate analysis methods, otherwise the model based on the data may not match the facts. Machine learning models represented by neural networks collect massive data to form decision models, which are helpful to accurately predict the water level in front of cascade pumping stations, improve the operating efficiency of pumping stations, and realize intelligent scheduling, which is of great significance to the construction of water networks. Nowadays, water level prediction is usually based on the construction of neural network models, such as using traditional neural network models such as BP and RNN, or using composite neural network models such as GA-BP and GRA-NARX. Through the water level prediction method driven by mechanism and data, it is discussed from different angles, but how to achieve real-time water level prediction is still a difficult problem that needs to be overcome. Real-time prediction of the water level in front of cascade pumping stations is conducive to ensuring water supply safety, optimizing decision-making and scheduling, and improving the efficiency of water resource utilization. Therefore, it is necessary to use the LSTM model with real-time prediction advantages to predict the water level in front of cascade pumping stations.

[0004] Since water level prediction faces the randomness and hysteresis of time series, traditional prediction methods usually use simple linear functions, but they often tend to ignore the influence of external conditions. In order to solve this problem, a real-time neural network prediction model is introduced. As a special recurrent neural network structure, LSTM has both the nonlinear characteristics of neural networks and memory and long-term dependence, which is convenient for its processing of time series data. However, while the recurrent neural network improves the prediction accuracy, it also prolongs the forecast period, which is very likely to lead to the accumulation of prediction errors. The LSTM real-time prediction model performs well in prediction accuracy and time series, but there is still no clear answer on how to solve the problems of real-time water monitoring data interference, time period prediction difficulties, cumbersome water level prediction input factors, and rolling accumulation of prediction errors. Therefore, technicians in this field provide a real-time prediction method for the water level in front of the cascade pump station with optimized input and rolling correction to solve the problems raised in the above background technology. Summary of the invention

[0005] 1. Technical issues to be solved

[0006] In view of the deficiencies in the prior art, the present invention provides a real-time prediction method for the water level in front of a cascade pump station with optimized input and rolling correction, which solves the problems of interference in real-time water situation monitoring data, difficulty in time period prediction, cumbersome water level prediction input factors, and rolling accumulation of prediction errors.

[0007] (II) Technical solution

[0008] To achieve the above purpose, the present invention is implemented by the following technical scheme: a real-time prediction method for the water level in front of a cascade pump station with optimized input and rolling correction, including pump station water level data, pump station flow data and two-level flow difference data, comprising the following steps:

[0009] S1. First, the collected water level, flow rate and two-stage flow difference data of the cascade pump station are cleaned;

[0010] S2. Perform normalization preprocessing on the cleaned data, and then use the Pearson correlation coefficient to calculate the correlation of the input factors. Only one input factor with a correlation coefficient greater than 0.6 is selected;

[0011] S3. After using correlation to preliminarily screen the input factors, perform sensitivity analysis on them using the OTA method to understand the impact of different input variables on model accuracy;

[0012] S4. Considering the reaction lag time of the channel section in the water flow propagation process, the dynamic wave theory and motion wave theory are used for calculation, and different forecast periods with the same time interval are selected for time-division forecasting;

[0013] S5. To prevent the error accumulation problem caused by the extension of the forecast period in the forecast model, error correction technology is introduced;

[0014] S6. Input the training set data of the water level in front of the station and the flow rate of the pump station into the LSTM-EC neural network model, and make the simulated water level value as close to the measured value as possible by continuously adjusting and optimizing the weight matrix and the bias term, so as to complete the training optimization of the LSTM-EC neural network model;

[0015] S7. Using the optimized model, input the prediction set data, and finally realize the real-time rolling correction prediction of the water level in front of the cascade pump station.

[0016] Preferably, in step S1, the water level and flow data of the water diversion project are arranged in a time series relationship to represent the working status of the pump station in different periods, and arranged in a space series relationship to represent the working status of the pump station in different periods. The arrangement in a space series relationship can reflect the water volume relationship of the current channel section. In order to ensure the logic and consistency of the flow data, it is necessary to clean the flow data at both time and space scales, and deduce the water loss rate of any two time periods according to the principle of dynamic balance of water volume:

[0017]

[0018] Divide the water loss flow by the inflow flow to obtain the dimensionless water loss rate:

[0019]

[0020] According to the longest sequence found, the flow data in the interval is cleaned using the principle of dynamic water balance:

[0021]

[0022] Preferably, in the step S2, in order to prevent the network gradient from disappearing and not converging, the input sequence needs to be standardized, and the input sequence is standardized to between 0 and 1 using the following formula:

[0023]

[0024] After normalization, the Pearson correlation coefficient was used to perform a preliminary screening of the input factors:

[0025]

[0026] Preferably, in the S3 step, the one-at-a-time method is used to perform parameter sensitivity analysis, that is, one input variable is changed, the baseline values ​​of other variables are maintained, and then the variables are returned to their standard values, and each input variable is repeated in a corresponding manner, thereby achieving an assessment of the sensitivity of the criterion weights.

[0027] Preferably, in step S4, considering the lag time in the water flow propagation process, before determining the forecast period, the dynamic wave theory and motion wave theory are used to calculate the channel section reaction lag time. The channel section reaction lag time calculated based on the dynamic wave theory can be expressed as:

[0028]

[0029] The formula for channel section reaction time calculated based on motion wave theory is:

[0030]

[0031] Calculation of Δτdw With Δτ kw , and the actual propagation time Δτ should be between the two, that is, Δτ dw <Δτ<Δτ kw , according to the actual propagation time, different forecast periods with the same time interval are selected for time period prediction, thereby greatly improving the accuracy of water level prediction.

[0032] Preferably, in the step S5, in order to prevent the error accumulation problem caused by the extension of the forecast period of the prediction model, an error correction technology is introduced to enable the prediction model to achieve real-time rolling correction over time, and continuously use real-time data as model input to update the future trend of the predicted water level, effectively prevent the error accumulation caused by long-term predictions, and improve the prediction accuracy while meeting the timeliness. The error autoregression correction method is based on the dependence law of the model prediction errors before and after, and uses the autocorrelation of the historical error series to predict future errors, and thereby correct the original prediction results of the model.

[0033] Preferably, in step S6, the LSTM-EC neural network structure adopts the following structure:

[0034] Forget Gate: F t =σ(X t W XF +H t-1 W HF +b F ), where W XF Represents the weight coefficient of the forget gate, H t-1 represents the input of the hidden layer at the previous moment, b F represents the forgetting bias, σ represents the sigmoid activation function

[0035] Input gate: I t =σ(X t W XI +H T-1 W HI +b I ), where W XI represents the input gate weight coefficient, b I Represents the bias of the input to the hidden layer;

[0036] Enter memory information: C t =i t *tanh(X t W XC +H t-1 W HC +b C )+f t *C t-1 , where W XC represents the memory unit weight, b CRepresents the bias from the input layer to the memory unit;

[0037] Output gate: O t =σ(X t W XO +H t-1 W HO +b O ), where W XO represents the output gate weight coefficient, b O Represents the deviation from the hidden layer to the output gate. The output of the hidden layer at time t is H t =O t *tanh(C t ).

[0038] (III) Beneficial effects

[0039] The invention provides a real-time prediction method for water level in front of a cascade pump station with optimized input and rolling correction.

[0040] It has the following beneficial effects:

[0041] 1. The present invention takes the multivariate LSTM neural network as the baseline model, and makes improvements on this basis. The relevant water level and flow data are cleaned before data preprocessing; the cleaned data is normalized, and the influence of different input variables on the model accuracy is mastered by increasing or decreasing the input factors, and the input factors with large sensitivity coefficients are focused on, so as to greatly reduce the complexity of the model, reduce the workload of data analysis and processing, and improve the accuracy of the model to a large extent, making up for the deficiency that the traditional LSTM neural network cannot perform sensitivity analysis due to the temporal nature of the LSTM neural network; before determining the forecast period, the dynamic wave theory and the motion wave theory are used to calculate the channel section reaction lag time, and different forecast periods with the same time interval are selected according to the actual propagation time for time-division forecasting, thereby greatly improving the water level forecasting accuracy; the error correction (EC) technology is introduced to enable the forecasting model to realize real-time rolling correction over time, and continuously use real-time data as model input to update the future trend of the predicted water level, effectively prevent the error accumulation caused by long-term forecasting, and improve the forecasting accuracy while meeting the timeliness. The present invention is expected to provide a new method and idea for the real-time forecasting of the water level in front of the cascade pump station, and the obtained results and laws can also provide data reference for the water level change trend of the front pool of the pump station. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a schematic diagram of the real-time prediction network of the water level in front of the LSTM-EC cascade pump station in the present invention;

[0043] Figure 2 This is a schematic diagram of the LSTM neural network model structure in the present invention;

[0044] Figure 3This is a schematic diagram of the real-time rolling correction forecast process of the present invention;

[0045] Figure 4 It is a schematic diagram of the prediction results of the water level in front of the cascade pumping station under different forecast periods by the LSTM-EC neural network model in the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] Embodiment 1:

[0048] like Figure 1-4 As shown, the embodiment of the present invention provides a real-time prediction method for the water level in front of a cascade pump station with optimized input and rolling correction, including pump station water level data, pump station flow data and two-level flow difference data, including the following steps:

[0049] S1. First, the collected water level, flow rate and two-stage flow difference data of the cascade pump station are cleaned;

[0050] S2. Perform normalization preprocessing on the cleaned data, and then use the Pearson correlation coefficient to calculate the correlation of the input factors. Only one input factor with a correlation coefficient greater than 0.6 is selected;

[0051] S3. After using correlation to preliminarily screen the input factors, perform sensitivity analysis on them through the OTA method to understand the influence of different input variables on the accuracy of the model, and focus on variables with large sensitivity coefficients, so as to greatly reduce the complexity of the model, reduce the workload of data analysis and processing, and greatly improve the accuracy of the model;

[0052] S4. Considering the reaction lag time of the channel section in the water flow propagation process, the dynamic wave theory and motion wave theory are used for calculation, and different forecast periods with the same time interval are selected for time-division forecasting;

[0053] S5. To prevent the error accumulation problem caused by the extension of the forecast period in the forecast model, error correction technology is introduced;

[0054] S6. Input the training set data of the water level in front of the station and the flow rate of the pumping station into the LSTM-EC neural network model, and make the water level simulation value as close to the measured value as possible by continuously adjusting and optimizing the weight matrix and the deviation term, and complete the training optimization of the LSTM-EC neural network model. The parameter settings of the LSTM-EC neural network real-time prediction model constructed by the present invention are as follows: the hidden layer size is 64, the batch value size is 500, the column length is 1, the loss function is MSE, and the step size is set according to the forecast period of 2h, 4h and 6h, and 500 iterations are performed; input the training set data of the pumping station water level, pumping station flow rate, etc. into the LSTM neural network model, and make the water level simulation value as close to the measured value as possible by continuously adjusting and optimizing the weight matrix and the deviation term, and complete the training optimization of the LSTM neural network model; use the optimized model, input the prediction set data, and finally realize the prediction of the water level in front of the station;

[0055] In order to further simulate the effectiveness of the method of the present invention, the experimental data of the present invention is that the water level and flow rate are collected by sensors at a pump station in a certain place, and corresponding standards are set.

[0056] S7. Using the optimized model, input the prediction set data, and finally realize the real-time rolling correction prediction of the water level in front of the cascade pump station.

[0057] In step S1, the water level and flow data of the water diversion project are arranged in a time series relationship to represent the working status of the pump station in different periods, and in a spatial series relationship to represent the working status of the pump station in different periods. The spatial series relationship can reflect the current water volume relationship of the canal section. In order to ensure the logic and consistency of the flow data, it is necessary to clean the flow data at both time and space scales, and deduce the water loss rate in any two time periods based on the principle of dynamic balance of water volume:

[0058]

[0059] Divide the water loss flow by the inflow flow to obtain the dimensionless water loss rate:

[0060]

[0061] According to the longest sequence found, the flow data in the interval is cleaned using the principle of dynamic water balance:

[0062]

[0063] In step S2, in order to prevent the network gradient from disappearing and not converging, the input sequence needs to be standardized. The following formula is used to standardize the input sequence to between 0 and 1:

[0064]

[0065] After normalization, the Pearson correlation coefficient was used to perform a preliminary screening of the input factors:

[0066]

[0067] In step S3, the one-at-a-time method is used to perform parameter sensitivity analysis, that is, changing one input variable, keeping the baseline values ​​of other variables, and then returning the variables to their standard values. This is repeated in the same way for each input variable, thereby achieving an assessment of the sensitivity of the criterion weights.

[0068] In step S4, considering the lag time in the water flow propagation process, the dynamic wave theory and motion wave theory are used to calculate the channel section reaction lag time before determining the forecast period. The channel section reaction lag time calculated based on the dynamic wave theory can be expressed as:

[0069]

[0070] The formula for channel section reaction time calculated based on motion wave theory is:

[0071]

[0072] Calculation of Δτ dw With Δτ kw , and the actual propagation time Δτ should be between the two, that is, Δτ dw <Δτ<Δτ kw , according to the actual propagation time, different forecast periods with the same time interval are selected for time period prediction, thereby greatly improving the accuracy of water level prediction.

[0073] In step S5, in order to prevent the error accumulation problem caused by the extension of the forecast period of the forecast model, the error correction technology is introduced to enable the forecast model to realize real-time rolling correction over time, and continuously use real-time data as the model input to update the future trend of the predicted water level, effectively prevent the error accumulation caused by long-term prediction, and improve the prediction accuracy while meeting the timeliness. The error autoregression correction method is based on the dependence law of the model prediction error before and after, and uses the autocorrelation of the historical error series to predict the future error, and correct the original prediction result of the model; the real-time rolling prediction idea is used to establish a real-time rolling correction prediction framework, and the LSTM is coupled with the EC technology. The real-time data is continuously used to update over time and used as the new input of the model for prediction, in order to improve the prediction effect of the water level in front of the cascade pump station hour by hour, and better meet the actual application requirements; assuming that the current time is T, it is necessary to make a rolling prediction of the water level in front of the cascade pump station, and set the forecast period of each rolling moment to 2h; the specific operation process is as follows:

[0074] Based on the historical water level data, a real-time water level prediction model based on LSTM-EC is constructed to predict the historical data, and the historical sequence prediction error is obtained by subtracting the historical actual value. Then, the autoregressive term at the input end of the error correction model is determined based on the autocorrelation analysis, so as to establish the error correction model.

[0075] Based on the existing data extraction, the input data required by the prediction model and the error autoregression correction model are calculated, and the water level and error results of the next forecast period are predicted on this basis. Then, the next forecast period is predicted based on the results of the next forecast period, and so on, and finally the water level prediction results and error prediction results within the next 6 hours are obtained;

[0076] Combine the above results to obtain the water level correction process in the next 6 hours;

[0077] Move the forecast period to T=T+1, use the latest actual water level data of the current forecast period to replace the model forecast value of the previous forecast period, calculate the actual forecast error and replace the error forecast value of the previous forecast period, then repeat the above forecast process, and finally obtain the new water level change process within the next 6 hours corresponding to the current rolling moment;

[0078] And so on, as the forecast period progresses, the above steps are repeated to continuously update the forecast results of the water level change process.

[0079] In step S6, the LSTM-EC neural network structure adopts the following structure:

[0080] Forget Gate: F t =σ(X t W XF +H t-1 W HF +b F ), where W XF Represents the weight coefficient of the forget gate, H t-1 represents the input of the hidden layer at the previous moment, b F represents the forgetting bias, σ represents the sigmoid activation function

[0081] Input gate: I t =σ(X t W XI +H t-1 W HI +b I ), where W XI represents the input gate weight coefficient, b I Represents the bias of the input to the hidden layer;

[0082] Enter memory information: C t =i t*tanh(X t W XC +H t-1 W HC +b C )+f t *C t-1 , where W XC represents the memory unit weight, b C Represents the bias from the input layer to the memory unit;

[0083] Output gate: O t =σ(X t W XO +H t-1 W HO +b O ), where W XO represents the output gate weight coefficient, b O Represents the deviation from the hidden layer to the output gate. The output of the hidden layer at time t is H t =O t *tanh(C t ).

[0084] This LSTM-EC cascade pump station front water level real-time prediction model uses mean absolute error MAE, mean square error MSE, root mean square error RMSE, mean absolute percentage error MAPE, correlation coefficient r and determination coefficient R2 as indicators to evaluate the model performance. The specific calculation formulas are:

[0085]

[0086] Among them, f i is the predicted value, y i is the measured value, m is the number of samples, is the average value before the test begins; MAE represents the average value of the absolute error between the predicted value and the measured value; MSE is the expected value of the square difference between the predicted value and the measured value; RMSE represents the sample standard deviation of the difference between the predicted value and the measured value; MAPE, as a relative error measure, uses the measured value to prevent positive and negative errors from offsetting each other; the correlation coefficient r is used to determine the degree of linear correlation between the predicted value and the measured value; the determination coefficient R2 is a statistic that measures the goodness of fit of the regression model;

[0087] The hardware environment used in this experiment is: 1ntol(R)Core(TM)i5-10500 CPU@3.10GHz processor, Nvidia GeForce GTX 1050Ti graphics card, 16G memory;

[0088] The software environment is: operating system Windows 11, acceleration environment CUDA 11.1, programming language Python 3.7, deep learning framework Pytorch 1.11; in the network model training phase, the hidden layer size is 64, the batch value size is 500, the column length is 1, the loss function is MSE, the step size is set according to the forecast period 2h, 4h and 6h, and 500 iterations are performed; the initial learning rate is 0.001, and the momentum factor is 0.9; as shown in the following table:

[0089]

[0090] Table 1 shows the performance evaluation indicators of the model under different forecast periods of 2h, 4h and 6h. The MAE ranges from 0.0173 to 0.0176. No matter how long the forecast period is, the average error of the model remains almost stable, indicating that the model has high stability in different forecast periods; the MSE ranges from 0.0006 to 0.0011, and gradually decreases with the increase of the forecast period, further verifying the improvement of the model's accuracy when dealing with longer time period forecasts; the RMSE ranges from 0.0257 to 0.0327, which is consistent with the trend of MSE. The smaller value indicates that the model is stable in the difference between the predicted value and the actual value; the MAPE value remains unchanged. It can be seen that the model shows high stability in different forecast periods, and the MAE, MSE, RMSE and MAPE indicators all show the consistency of the forecast performance in each forecast period; in summary, the LSTM-EC neural network real-time prediction rolling correction model shows good prediction accuracy, high stability and accurate long-term trend capture ability in the time period prediction task spanning multiple forecast periods, which fully proves its reliability in dealing with complex time series time period prediction problems and the feasibility of real-time rolling correction of prediction results, and can realize accurate prediction of the water level in front of cascade pumping stations.

[0091] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time prediction method for water level in front of a cascade pump station with optimized input and rolling correction, including pump station water level data, pump station flow data and two-stage flow difference data, characterized in that: The following steps are involved: S1. First, the collected water level, flow rate and two-stage flow difference data of the cascade pump station are cleaned; S2. Perform normalization preprocessing on the cleaned data, and then use the Pearson correlation coefficient to calculate the correlation of the input factors. Only one input factor with a correlation coefficient greater than 0.6 is selected; S3. After using correlation to preliminarily screen the input factors, perform sensitivity analysis on them using the OTA method to understand the impact of different input variables on model accuracy; S4. Considering the reaction lag time of the channel section in the water flow propagation process, the dynamic wave theory and motion wave theory are used for calculation, and different forecast periods with the same time interval are selected for time period prediction; S5. To prevent the error accumulation problem caused by the extension of the forecast period in the forecast model, error correction technology is introduced; S6. Input the training set data of the water level in front of the station and the flow rate of the pump station into the LSTM-EC neural network model, and make the simulated water level value as close to the measured value as possible by continuously adjusting and optimizing the weight matrix and the bias term, so as to complete the training optimization of the LSTM-EC neural network model; S7. Using the optimized model, input the prediction set data, and finally realize the real-time rolling correction prediction of the water level in front of the cascade pump station.

2. The method for real-time prediction of water level in front of a cascade pump station with optimized input and rolling correction according to claim 1 is characterized by: In step S1, the water level and flow data of the water diversion project are arranged in a time series relationship to represent the working status of the pump station in different periods, and arranged in a spatial series relationship to represent the working status of the pump station in different periods. The arrangement in a spatial series relationship can reflect the water volume relationship of the current channel section. In order to ensure the logic and consistency of the flow data, it is necessary to clean the flow data at both time and space scales, and deduce the water loss rate of any two time periods based on the principle of dynamic balance of water volume: Divide the water loss flow by the inflow flow to obtain the dimensionless water loss rate: According to the longest sequence found, the flow data in the interval is cleaned using the principle of dynamic water balance:

3. The method for real-time prediction of water level in front of cascade pumping stations with optimized input and rolling correction according to claim 1 is characterized by: In the S2 step, in order to prevent the network gradient from disappearing and not converging, the input sequence needs to be standardized. The following formula is used to standardize the input sequence to between 0 and 1: After normalization, the Pearson correlation coefficient was used to perform a preliminary screening of the input factors:

4. The method for real-time prediction of water level in front of a cascade pump station with optimized input and rolling correction according to claim 1 is characterized by: In the S3 step, the one-at-a-time method is used to perform parameter sensitivity analysis, that is, changing one input variable, keeping the baseline values ​​of other variables, and then returning the variables to their standard values, and repeating for each input variable in the desired manner, thereby achieving the evaluation of the sensitivity of the criterion weights.

5. The method for real-time prediction of water level in front of a cascade pumping station with optimized input and rolling correction according to claim 1 is characterized by: In step S4, considering the lag time in the water flow propagation process, before determining the forecast period, the dynamic wave theory and motion wave theory are used to calculate the channel section reaction lag time. The channel section reaction lag time calculated based on the dynamic wave theory can be expressed as: The formula for channel section reaction time calculated based on motion wave theory is: Calculation of Δτ dw With Δτ kw , and the actual propagation time Δτ should be between the two, that is, Δτ dw <Δτ<Δτ kw , according to the actual propagation time, different forecast periods with the same time interval are selected for time period prediction, thereby greatly improving the accuracy of water level prediction.

6. The method for real-time prediction of water level in front of a cascade pump station with optimized input and rolling correction according to claim 1 is characterized by: In the step S5, in order to prevent the error accumulation problem caused by the extension of the forecast period of the forecast model, an error correction technology is introduced to enable the forecast model to achieve real-time rolling correction over time, and continuously use real-time data as model input to update the future trend of the predicted water level, effectively preventing the error accumulation caused by long-term forecasts, and improving the forecast accuracy while meeting the timeliness. The error autoregression correction method is based on the dependence law of the model prediction errors before and after, and uses the autocorrelation of the historical error series to predict future errors, and thereby correct the original prediction results of the model.

7. The method for real-time prediction of water level in front of a cascade pump station with optimized input and rolling correction according to claim 1 is characterized by: In step S6, the LSTM-EC neural network structure adopts the following structure: Forget Gate: F t =σ(X t W XF +H t-1 W HF +b F ), where W XF Represents the weight coefficient of the forget gate, H t-1 represents the input of the hidden layer at the previous moment, b f represents the forgetting bias, σ represents the sigmoid activation function Input gate: I t =σ(X t W XI +H t-1 W HI +b I ), where W Xl represents the input gate weight coefficient, b l Represents the bias of the input to the hidden layer; Enter memory information: C t =i t *tanh(X t W XC +H t-1 W HC +b C )+f t *C t-1 , where W XC represents the memory unit weight, b C Represents the bias from the input layer to the memory unit; Output gate: O t =σ(X t W XO +H t-1 W HO +b O ), where W XO represents the output gate weight coefficient, b O Represents the deviation from the hidden layer to the output gate. The output of the hidden layer at time t is H t =O t *tanh(C t ).

Images