Gate pump station water level prediction method based on microwave link rain measurement data

Abstract

The invention provides a gate pump station water level prediction method based on microwave link rain measurement data, and belongs to the field of microwave signal processing and hydrological application. The method comprises the steps of collecting data of a gate pump station; utilizing the collected microwave link attenuation time sequence data to invert regional rainfall intensity; processing the inverted regional rainfall intensity together with the upstream and downstream water levels of the gate pump station, the gate opening height and the pump operation state into multi-input training data; constructing a multi-input single-output water level prediction model; training the water level prediction model by using the training data; and using the trained water level prediction model to predict the water level of the gate pump station. According to the method, key factors influencing the water level change of the gate pump station are fully considered, a water level prediction model based on LSTM + Attention is designed, and accurate prediction of the upstream water level of the gate pump station can be achieved.

Description

Technical Field

[0001] The present invention relates to the fields of microwave signal processing and hydrological applications, and particularly to a method for predicting the water level of a sluice pump station based on rainfall data measured by a microwave link. Background Art

[0002] As an important water conservancy project hub, the sluice pump station plays an important role in flood control, drainage, water diversion, water environment treatment and other aspects. Accurate water level prediction can provide a scientific basis for the operation and dispatching of the sluice pump station, which is conducive to managers predicting water regime changes and thus making reasonable decisions. The water level prediction of the sluice pump station not only helps to prevent and mitigate floods and improve the ability to respond to flood disaster risks, but also is crucial for ensuring water supply safety, maintaining the ecological balance of rivers and water areas, and promoting the sustainable utilization of water resources.

[0003] Rainfall is a key factor causing water level fluctuations. The error of rainfall information will be transmitted to the water level prediction value through the hydrological model, resulting in an increase in the uncertainty of water level prediction. Therefore, improving the accuracy of rainfall information is the key to ensuring accurate water level prediction. Commercial microwave links (CML) in the communication network, as a new rainfall measurement method, have certain advantages in terms of spatial resolution, monitoring accuracy and real-time performance compared with current traditional rainfall monitoring means, such as rain gauges, radars, satellites, etc., and can provide a more accurate rainfall data source for the water level prediction model. On the other hand, the water level prediction based on traditional physical models has high requirements for terrain and underground pipe network data. Especially in urban areas, road construction, pipeline maintenance, etc. will cause difficulties in collecting basic geographical data, and the data quality cannot be guaranteed. Summary of the Invention

[0004] In view of this, the present invention proposes a method for predicting the water level of a sluice pump station based on rainfall data measured by a microwave link. The present invention uses the rainfall and multi-modal working condition data inversed from the CML attenuation time series, and the LSTM model with an Attention mechanism to realize the prediction of the upstream water level of the sluice pump station, providing a scientific basis for the operation and dispatching of water conservancy projects and managers to make decisions.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A method for predicting the water level of a sluice pump station based on rainfall data measured by a microwave link, comprising the following steps:

[0007] Step 1, collecting data of the sluice pump station, including the microwave link attenuation time series, the upstream and downstream water levels of the sluice pump station, the opening height of the gate and the pump operation status;

[0008] Step 2, inversing the regional rainfall intensity by using the collected microwave link attenuation time series data;

[0009] Step 3: Process the inverted regional rainfall intensity together with the water levels upstream and downstream of the sluice pumping station, the opening height of the gate, and the pump operating status into multi-input training data;

[0010] Step 4: Construct a multi-input single-output water level prediction model;

[0011] Step 5: Use the training data to train the water level prediction model;

[0012] Step 6: Use the trained water level prediction model to predict the water level of the sluice pumping station.

[0013] Furthermore, the specific method of Step 2 is as follows:

[0014] Step 201: Calculate the microwave signal attenuation time series A(t):

[0015] A(t) = TSL - RSL

[0016] where TSL is the transmitted signal level and RSL is the received signal level;

[0017] Step 202: Calculate the moving standard deviation Std(t) of A(t):

[0018]

[0019] where W is the moving window length, N W is the total number of samples within the window, and the overline represents the average value;

[0020] Set a threshold q. If Std(t) > q, then t is the rainfall period; otherwise, t is the non-rainfall period;

[0021] Step 203: Determine the baseline attenuation A b (t): During the non-rainfall period, the baseline attenuation is equal to the current attenuation value; during the rainfall period, the baseline attenuation is the sum of the attenuation value of the nearest non-rainfall period before the current moment and the attenuation value of the wet antenna;

[0022] Step 204: Subtract the baseline attenuation from the original microwave link attenuation time series to obtain the microwave rain-induced attenuation A rain (t):

[0023] A rain (t) = A(t) - A b (t)

[0024] Step 205: Invert the path-average rainfall intensity R through the ITU attenuation rate - rainfall rate relationship;

[0025]

[0026] Among them, k and α are coefficients related to the microwave carrier frequency, link length, and raindrop spectrum, and L is the link path length.

[0027] Further, the specific method of step 3 is as follows:

[0028] Duplicate value detection and processing: Delete all duplicate rows and keep only one of them.

[0029] Outlier filtering: Delete outliers that exceed the set range of water level monitoring and whose gate opening height is less than zero from the time series data.

[0030] Time interval unification: Set the time as the index and re-index according to the sampling frequencies of various types of data to make the time intervals of each type of data consistent and uniform.

[0031] Missing value processing: Perform linear interpolation on the missing values corresponding to the new index.

[0032] Timestamp alignment: Adjust and align various types of data with different time resolutions, that is, merge the 1-minute time resolution of microwave link rain gauge data into one sample every 5 minutes.

[0033] Integrate all input data into a document by column according to the timestamp to form multi-input training data.

[0034] Further, in step 4, a water level prediction model is constructed based on the LSTM long short-term memory network and the Attention attention mechanism. The working method of the water level prediction model is as follows:

[0035] Use the input feature x at the current moment t t and the hidden state h at the previous moment t-1 of the LSTM t-1 to calculate the forget gate f t , input gate i t , output gate o t , and candidate value z t :

[0036] f t = σ(W f x t + U f h t-1 + b f )

[0037] i t = σ(W i x t + U i h t-1 + b i )

[0038] o t = σ(W 0 xt +U 0 h t-1 +b 0 )

[0039] z t = tanh(W z x t +U z h t-1 +b z )

[0040] Wherein, σ(·) and tanh(·) are the Sigmoid function and the hyperbolic tangent function respectively; the weight matrices W f , W i , W z , W o , U f , U i , U z , U o and the bias vectors b f , b i , b z , b o are obtained during the training process;

[0041] Combined with f t , i t , z t and the memory cell state c t-1 at the previous moment, update the memory cell state c t at the current moment:

[0042] c t = i t × z t + f t × c t-1

[0043] Then, pass the information of c t to the hidden state h t at the current moment through o t :

[0044] h t = o t × tanh(c t )

[0045] Next, using the output h tAs the input of the Attention layer, create a weight matrix and a bias term, multiply the input by the weight matrix, add the bias term b, then use the tanh activation function to perform a non-linear transformation on the result to obtain the attention scores, and then use the softmax function to normalize the attention scores. Multiply the attention weights by the input to obtain the weighted features at each time step;

[0046] Finally, output the target value y through a fully connected layer pred_norm , and perform denormalization on the target value to obtain the water level prediction value y pred :

[0047] y pred = y pred_norm × std + mean;

[0048] where mean is the mean of the input data and std is the standard deviation of the input data.

[0049] Furthermore, in step 5, adopt batch training and sequence-to-sequence training methods to train the water level prediction model batch by batch. Use all data in the past 5 hours as the input sequence and the water level in the next 1 hour as the output sequence; the specific process is as follows:

[0050] Add a sliding window to the training data, with a sliding step of 1 data point and a window length set to 6 hours. All features in the first 5 hours within each window are used as the input to predict the target value 1 hour later;

[0051] The input data undergoes forward propagation through the hidden layer to calculate the predicted value;

[0052] Compare the predicted value with the true value and calculate the mean squared error loss function value;

[0053] According to the loss function, calculate the gradient of each weight in the network by backpropagating the error through time;

[0054] Use the Adam optimization algorithm to update the weights of the model according to the calculated gradients;

[0055] Stop training after reaching the set number of iterations.

[0056] Furthermore, the specific method in step 6 is as follows:

[0057] Obtain the real-time attenuation of microwave links, the measured water levels upstream and downstream of the gates, and all gate opening height data within 5 hours before the current moment in the region; if it is a combined gate-pump station, the input data also includes all pump operating status information;

[0058] Invert the regional rainfall intensity through the real-time attenuation data of the microwave link, preprocess and normalize the input data, and then input it into the trained water level prediction model. The output of the model is the predicted value of the upstream water level of the sluice pump station for the next 1 hour at the current moment.

[0059] The beneficial effects of the present invention are as follows:

[0060] 1. The present invention adopts a multi-modal time series data processing algorithm and constructs a water level prediction model based on the long short-term memory (LSTM) and attention mechanisms, which can realize the water level prediction of the sluice pump station based on CML rainfall measurement data.

[0061] 2. For multi-modal time series data such as microwave attenuation, upstream and downstream water levels, gate opening heights, and pump station operating states, the present invention preprocesses the input data by using methods such as sliding window segmentation, dynamic time warping, and standard normalization. In addition, the present invention selects LSTM, which is very suitable for processing time series data, to automatically learn the long-term dependencies in the sequence, and adds a Soft Attention mechanism between the LSTM model and the fully connected layer to assign higher weights to key time steps, which helps the model capture more complex and subtle patterns.

[0062] 3. The hybrid model of the present invention takes CML rainfall measurement data and multi-modal engineering situation data as inputs and outputs the predicted value of the upstream water level of the sluice pump station with a certain prediction period, thereby providing a basis and guidance for ensuring the safe operation of water conservancy projects and formulating emergency management decisions. Description of the Drawings

[0063] Figure 1 is a flowchart of the method for predicting the water level of the sluice pump station based on CML rainfall measurement data in the present invention.

[0064] Figure 2 is an effect diagram of rainfall inversion using the CML received signal level.

[0065] Figure 3 is a schematic diagram of the water level prediction model based on LSTM+Attention.

[0066] Figure 4 is an effect diagram comparing the predicted water level value of the model with the actual measured water level value. Detailed Embodiments

[0067] The following further elaborates on the present invention with reference to the drawings.

[0068] A method for predicting the water level of a sluice pumping station based on microwave link rainfall data. First, collect data closely related to the prediction target, i.e., the water level upstream of the sluice pumping station, including climate environmental factors and human interference factors. Second, use the collected CML attenuation data to invert the regional rainfall intensity. Third, construct a deep learning model capable of processing multi-modal time series inputs and initialize the model parameters. Finally, divide the training set and test set, sample the input features and target values, and optimize the model parameters by minimizing the loss function. As Figure 1 shown, the specific steps are as follows:

[0069] First, collect the CML attenuation time series, the water levels upstream and downstream of the sluice pumping station, the opening height of the gate, and the pump operation status data, and use the CML attenuation time series for rainfall inversion. Assume that the transmitted and received signal levels are TSL (dBm) and RSL (dBm) respectively. Then, the microwave signal attenuation time series A(t) (dB) can be obtained by calculating A(t)=TSL - RSL. Calculate the moving standard deviation Std(t) of A(t):

[0070]

[0071] where W is the moving window length and N W is the total number of samples within the window. If Std(t)>q, t is the rainfall period; otherwise, it is the non-rainfall period. q is a threshold related to the ratio of rainfall time to the total observation time.

[0072] Determine the baseline attenuation A b (t): During the non-rainfall period, the baseline is equal to the current attenuation value, i.e., A b (t)=A(t); during the rainfall period, the baseline value is the sum of the attenuation value of the nearest non-rainfall period before the current moment and the wet antenna attenuation value, i.e., A b (t)=A lastdry (t)+A WAA .

[0073] Subtract the baseline from the original CML attenuation time series to obtain the microwave rain-induced attenuation, i.e.: A rain (t)=A(t)-A b (t). Finally, invert the path-averaged rainfall intensity R (mm / h) through the ITU attenuation rate - rainfall rate relationship:

[0074]

[0075] where k and α are coefficients related to the microwave carrier frequency, link length, and raindrop size distribution, and L (km) is the link path length. The CML received signal level time series and the rainfall intensity curve obtained by inversion are as Figure 2 shown.

[0076] Subsequently, perform operations such as missing value processing and timestamp alignment on all input feature data, and divide the overall data into a training set and a test set at a ratio of 4:1. Perform standard normalization using the mean and variance of the training set. Then, sample the input features and target values in the form of a sliding window on the training and test data sets. For example, use all the features in the first 5 hours within a window as the input to predict the target value (upstream water level) 1 hour later.

[0077] Next, construct a water level prediction model based on LSTM+Attention, and use the current input feature x t and the hidden state h t-1 of the LSTM model at the previous moment to calculate the forget gate f t , input gate i t , output gate o t , and candidate value z t :

[0078] f t = σ(W f x t + U f h t-1 + b f )

[0079] i t = σ(W i x t + U i h t-1 + b i )

[0080] o t = σ(W 0 x t + U 0 h t-1 + b 0 )

[0081] z t = tanh(W z x t + U z h t-1 + b z )

[0082] In the formula, σ(·) and tanh(·) are the Sigmoid function and the hyperbolic tangent function respectively; the weight matrices W f , W i , W z , W o , U f , U i , U z , U o and the bias vectors b f , bi , b z , b o can be obtained during the training process. Combining f t , i t , z t and the memory cell state c at the previous moment t-1 to update the memory cell state c at the current moment t ; Pass c t information to the hidden state h at the current moment t : t :

[0083] c t = i t × z t + f t × c t-1

[0084] h t = o t × tanh(c t )

[0085] Then, using the output h of the LSTM t as the input of the Attention layer, create the matrix weights and bias terms, multiply the input input by the weight matrix, and add the bias term b. Then use the tanh activation function to perform a non-linear transformation on the result to obtain the attention scores, and then use the softmax function to normalize the attention scores. Multiply the attention weights by the input to obtain the weighted features for each time step. Finally, output the target value through the fully connected layer and perform anti-normalization on it to obtain the water level prediction value. The entire model structure is as Figure 3 shown.

[0086] Finally, by comparing the predicted water level value of the model with the actual collected water level value, a scatter plot is drawn to view the prediction results. The root mean square error (RMSE) and Pearson correlation coefficient (PCC) between the two can be further calculated to comprehensively evaluate the model. Different input features and target value lengths can also be modified to evaluate the maximum prediction period range of the model. The comparison effect between the predicted water level value and the actual water level value is as Figure 4 shown.

[0087] This method fully considers the key factors affecting the water level change of the sluice pump station, collects relevant basic data, and processes these multi-modal time series data to meet the model input requirements. On this basis, a water level prediction model based on LSTM + Attention is designed and constructed. Through training and testing, accurate prediction of the upstream water level of the sluice pump station is achieved. Finally, it is compared with the actual collected water level value to comprehensively evaluate the model performance.

[0088] The principle of this method is as follows:

[0089] Data collection: Combining the actual water area environment and application scenarios, analyze the factors affecting the change of the prediction target, and collect data related to the water levels of the sluice pumping stations, mainly including the CML attenuation time series (dBm), the upstream and downstream water levels of the sluice pumping stations (m), the opening height of the gate (m), and the pump operation status (0 represents closed, 1 represents open).

[0090] Rainfall inversion: Perform sunny / rainy discrimination, baseline attenuation determination, wet antenna attenuation compensation, and regional rainfall intensity inversion on the collected CML attenuation time series.

[0091] Multi-modal time series data processing: Check the quality of all input data, filter outliers, unify the time resolution, linearly interpolate short-term missing values, and align the timestamps of various data.

[0092] Prediction model construction: Construct an LSTM+Attention multi-input single-output time series prediction model and initialize the model parameters, mainly including the number of hidden layers, batch size, learning rate, regularization method, etc.

[0093] Training and testing: Divide the overall data into training set and test set, perform data normalization, and then sample the input features and target values in the form of a sliding window, determine the number of training epochs, start training and testing, and adjust and optimize the model parameters with the goal of reducing the loss.

[0094] Model verification: For the test set, compare the predicted water level values of the model with the measured water level values of the sensors, calculate RMSE and PCC, and evaluate the model performance by comparing the water level prediction results under different lead times.

[0095] Application scenario analysis: According to the water level prediction situation, formulate the opening duration and height of the regulating gate in advance or with a delay, the operation time and number of pumping stations, etc.

[0096] The following is a more specific example:

[0097] A method for predicting the water levels of sluice pumping stations based on microwave link rainfall measurement data, including the following steps:

[0098] Step 1, collect data of the sluice pumping stations, including the microwave link attenuation time series, the upstream and downstream water levels of the sluice pumping stations, the opening height of the gate, and the pump operation status;

[0099] Step 2, use the collected microwave link attenuation time series data to invert the regional rainfall intensity; the specific method is:

[0100] Step 201, calculate the microwave signal attenuation time series A(t):

[0101] A(t) = TSL - RSL

[0102] Among them, TSL is the transmitted signal level, and RSL is the received signal level;

[0103] Step 202, calculate the sliding standard deviation Std(t) of A(t):

[0104]

[0105] Among them, W is the sliding window length, and N W is the total number of samples within the window, and the overline represents the average value;

[0106] Set a threshold q. If Std(t) > q, then t is the rainfall period; otherwise, t is the non-rainfall period;

[0107] Step 203, determine the baseline attenuation A b (t): During the non-rainfall period, the baseline attenuation is equal to the current attenuation value; during the rainfall period, the baseline attenuation is the sum of the attenuation value in the nearest non-rainfall period before the current moment and the attenuation value of the wet antenna;

[0108] Step 204, subtract the baseline attenuation from the original CML attenuation time series to obtain the microwave rain-induced attenuation A rain (t):

[0109] A rain (t) = A(t) - A b (t)

[0110] Step 205, inversely calculate the path-average rainfall intensity R through the ITU attenuation rate - rainfall rate relationship:

[0111]

[0112] Among them, k and α are coefficients related to the microwave carrier frequency, link length, and raindrop size distribution, and L is the link path length.

[0113] Step 3, process the inversely calculated regional rainfall intensity together with the water levels upstream and downstream of the sluice pumping station, the opening height of the gate, and the pump operating status into multi-input training data; the specific method is as follows:

[0114] Duplicate value detection and processing: Delete all duplicate rows and only keep one of them;

[0115] Outlier filtering: Delete outliers that exceed the set range of water level monitoring (2 meters - 6 meters) and outliers with a gate opening height less than zero from the time series data;

[0116] Time interval unification: Set the time as the index and re-index according to the sampling frequencies of various types of data (CML rain measurement data is usually 1 minute; water level, gate, and pump data is 5 minutes) to make the time intervals of each type of data consistent and uniform;

[0117] Missing value handling: Perform linear interpolation on the missing values corresponding to the new index, that is, take the average of the values at the moments before and after the missing value;

[0118] Timestamp alignment: Adjust and align various data with different time resolutions, that is, merge the 1-minute time resolution of CML rainfall data into a sample once every 5 minutes;

[0119] Integrate all input data into a document column by column according to the timestamps.

[0120] Step 4, construct a multi-input single-output water level prediction model based on the LSTM long short-term memory network and the Attention attention mechanism; divide all the data into a training set and a test set in a ratio of 4:1. Perform Z-score standardization on the training set and test set data respectively with the mean and standard deviation of the training set:

[0121]

[0122] Z train_data =(X train_data (i)-mean) / std

[0123] Z test_data =(X test_data (i)-mean) / std

[0124] X represents the original data value, Z represents the normalized data, and N is the total number of samples.

[0125] Add sliding windows to the training set and test set respectively. The sliding step size is 1 data point, and the window length is set to 6 hours. All features (rainfall, upstream water level, downstream water level, gate opening height, pump station operation status) in the first 5 hours within each window are used as inputs to predict the target value (upstream water level) 1 hour later.

[0126] Use the input feature x at the current moment t t and the hidden state h of the LSTM at the previous moment t-1 t-1 , and calculate the forget gate f t , input gate i t , output gate o t , candidate value z t :

[0127] f t =σ(W f x t +U f h t-1 +b f )

[0128] i t =σ(W ix t +U i h t-1 +b i )

[0129] o t =σ(W 0 x t +U 0 h t-1 +b 0 )

[0130] z t =tanh(W z x t +U z h t-1 +b z )

[0131] where σ(·) and tanh(·) are the Sigmoid function and the hyperbolic tangent function respectively; the weight matrices W f 、W i 、W z 、W o 、U f 、U i 、U z 、U o and the bias vectors b f 、b i 、b z 、b o are obtained during the training process;

[0132] Combining f t 、i t 、z t and the memory cell state c t-1 at the previous moment, update the memory cell state c t at the current moment:

[0133] c t =i t ×z t +f t ×c t-1

[0134] Then, pass the information of c t to the hidden state h t at the current moment through o t :

[0135] h t =o t ×tanh(c t )

[0136] Next, using the output h tAs the input of the Attention layer, create a weight matrix and bias term, multiply the input by the weight matrix, add the bias term b, and then use the tanh activation function to perform a nonlinear transformation on the result to obtain the attention score. Then use the softmax function to normalize the attention score, multiply the attention weight by the input, and obtain the weighted feature of each time step;

[0137] Finally, the target value is output through the fully connected layer and denormalized to obtain the water level prediction value:

[0138] y pred =y pred_norm ×std+mean.

[0139] Step 5: Use the training data to train the water level prediction model; the specific method is:

[0140] Batch training and sequence-to-sequence training are used. The batch size is set to 256, and the network is trained batch by batch. All data in the past 5 hours are used as the input sequence, and the water level in the next hour is used as the output sequence.

[0141] The input data is forward propagated through the hidden layer to calculate the predicted value;

[0142] Compare the predicted value with the true value and calculate the mean square error loss function value;

[0143] According to the loss function, the error is back-propagated through time to calculate the gradient of each weight in the network;

[0144] Use the Adam optimization algorithm, set the learning rate to 0.001, and update the model weights based on the calculated gradients;

[0145] The initial number of iterations is set to 100, and the training stops when the set number of iterations is reached. At the same time, the training error, test error, and convergence of each iteration are observed.

[0146] Model evaluation: For the test set, the model predicts the water level value (y pred ) and the water level value measured by the water level gauge (y true ) to compare and calculate the root mean square error (RMSE) and Pearson correlation coefficient (PCC):

[0147]

[0148] And change the input features and target value length to evaluate the maximum forecast period of the model.

[0149] Step 6, using the trained water level prediction model to predict the water level of the sluice pump station;

[0150] The trained model can be applied to the water level prediction of the check gate and the site with a combination of check gates and pumps, and provide high-quality regional rainfall data for it. The necessary input data for the model includes the CML real-time attenuation within 5 hours before the current moment in the region, the measured water levels upstream and downstream of the gate, and all the gate opening height data. If it is a combined check gate and pump station, the input data also needs to add all the pump operation status information.

[0151] Invert the rainfall process using CML as described in step 2; preprocess the input data as described in step 3, and then it can be directly input into the model after Z-score normalization. The final output of the model is the predicted value of the water level upstream of the check gate and pump station for the next 1 hour at the current moment.

[0152] If you want to achieve a longer prediction period, such as water level predictions of 2 hours, 12 hours, or 24 hours, without changing the model structure, you can retrain the model parameters by adjusting the sliding window length described in step 4.

[0153] To sum up, the present invention uses the CML attenuation time series to invert the regional rainfall, combines it with multi-modal working condition data as feature inputs, and constructs an LSTM+Attention water level prediction model. Through training and testing, the predicted value of the water level upstream of the check gate and pump station is output, and the overall performance of the model is evaluated by comparing it with the real-time collected water level value. The present invention uses advanced microwave link rain measurement technology and deep learning models to achieve water level prediction of check gate and pump stations, which is of great significance for flood control and disaster reduction in urban watershed areas, improving the ability to respond to flood disaster risks, ensuring water supply safety, and maintaining the ecological balance of rivers and waters.

Claims

1. A method for predicting water level at a sluice pump station based on microwave link rainfall measurement data, characterized in that: The following steps are involved: Step 1, collect data of the gate pump station, including microwave link attenuation timing, upstream and downstream water levels of the gate pump station, gate opening height and pump operation status; Step 2, using the collected microwave link attenuation time series data to invert the regional rainfall intensity; Step 3, the inverted regional rainfall intensity together with the upstream and downstream water levels of the sluice pump station, the gate opening height and the pump operation status are processed into multi-input training data; Step 4, construct a water level prediction model with multiple inputs and single output; Step 5, using the training data to train the water level prediction model; Step 6: Use the trained water level prediction model to predict the water level of the sluice pump station.

2. The method for predicting water level of a sluice pump station based on microwave link rainfall data according to claim 1 is characterized in that: The specific method of step 2 is: Step 201, calculate the microwave signal attenuation time series A(t): A(t)=TSL-RSL Among them, TSL is the sending signal level, and RSL is the receiving signal level; Step 202, calculate the sliding standard deviation Std(t) of A(t): Where W is the sliding window length, N W is the total number of samples in the window, and the overline indicates the average value; Set the threshold q. If Std(t)>q, then t is a rainy period, otherwise t is a rainless period. Step 203, determine the baseline attenuation A b (t): In rainless periods, the baseline attenuation is equal to the current attenuation value; in rainy periods, the baseline attenuation is the sum of the attenuation value of the most recent rainless period before the current moment and the wet antenna attenuation value; Step 204: Subtract the baseline attenuation from the original microwave link attenuation time series to obtain the microwave rain-induced attenuation A rain (t): A rain (t)=A(t)-A b (t) Step 205, invert the path average rainfall intensity R through the ITU attenuation rate-rain rate relationship: Among them, k and α are coefficients related to the microwave carrier frequency, link length and raindrop spectrum, and L is the link path length.

3. The method for predicting water level of a sluice pump station based on microwave link rainfall data according to claim 2 is characterized in that: The specific method of step 3 is: Duplicate value detection and processing: delete all duplicate rows and keep only one row; Outlier filtering: outliers that exceed the water level monitoring setting range and gate opening height less than zero are deleted from the time series data; Unified time intervals: Set time as index, and re-index according to the sampling frequency of each type of data to make the time interval of each type of data consistent and uniform; Missing value processing: linear interpolation of missing values ​​corresponding to the new index; Timestamp alignment: adjust and align various data with different time resolutions, that is, merge the 1-minute time resolution of microwave link rain measurement data into a 5-minute sample; All input data are integrated into one document by column according to the timestamp to form multi-input training data.

4. The method for predicting water level of a sluice pump station based on microwave link rainfall data according to claim 3 is characterized in that: In step 4, a water level prediction model is built based on the LSTM long short-term memory network and the Attention mechanism. The working method of the water level prediction model is as follows: Using the input feature x at the current time t t and the hidden state h of LSTM at the previous time t-1 t-1 , calculate the forget gate f t , input gate i t , output gate o t , candidate value z t : f t =σ(W f x t +U f h t-1 +b f ) i t =σ(W i x t +U i h t-1 +b i ) the t =σ(W0x t +U0h t-1 +b0) z t =tanh(W z x t +U z h t-1 +b z ) Where σ(·) and tanh(·) are the sigmoid function and the hyperbolic tangent function respectively; the weight matrix W f , W i , W z , W o , U f , U i , U z , U o and the bias vector b f , b i , b z , b o Obtained during training; Combined with t 、i t 、z t And the memory cell state c at the previous moment t-1 , update the current memory cell state c t : c t =i t ×z t +f t ×c t-1 Then, through o t c t The information is passed to the current hidden state h t : h t =o t ×tanh(c t ) Next, take the LSTM output h t As the input of the Attention layer, create a weight matrix and bias term, multiply the input by the weight matrix, add the bias term b, and then use the tanh activation function to perform a nonlinear transformation on the result to obtain the attention score. Then use the softmax function to normalize the attention score, multiply the attention weight by the input, and obtain the weighted feature of each time step; Finally, the target value y is output through the fully connected layer pred_norm , and denormalize the target value to obtain the water level prediction value y pred : y pred =y pred_norm ×std+mean; Among them, mean is the mean of the input data, and std is the standard deviation of the input data.

5. The method for predicting water level of a sluice pump station based on microwave link rainfall measurement data according to claim 4 is characterized in that: In step 5, batch training and sequence-to-sequence training are used to train the water level prediction model batch by batch, taking all data from the past 5 hours as the input sequence and the water level in the next hour as the output sequence; the specific process is as follows: Add a sliding window to the training data, with a sliding step of 1 data point and a window length of 6 hours. All features of the first 5 hours in each window are used as input, and the target value is predicted 1 hour later. The input data is forward propagated through the hidden layer to calculate the predicted value; Compare the predicted value with the true value and calculate the mean square error loss function value; According to the loss function, the error is back-propagated through time to calculate the gradient of each weight in the network; Use the Adam optimization algorithm to update the model weights based on the calculated gradients; The training stops after reaching the set number of iterations.

6. A method for predicting water level at a sluice pump station based on microwave link rainfall data according to claim 5, characterized in that: The specific method of step 6 is: Obtain the real-time attenuation of the microwave link within 5 hours before the current time in the area, the measured water levels upstream and downstream of the gate, and the opening height data of all gates; if it is a gate-pump combination station, the input data also includes the operating status information of all pumps; The regional rainfall intensity is inverted through the real-time attenuation data of the microwave link, and the input data is preprocessed and normalized, and then input into the trained water level prediction model. The model output is the predicted value of the upstream water level of the sluice pump station in the next hour from the current moment.

Images