A prediction method for particulate matter deposition in drainage pipe networks based on long short-term memory step sequences

Abstract

The present invention provides a method for predicting particulate matter deposition in a drainage pipe network based on a long short-term memory step sequence, selects the main influencing factors and decision variables for predicting particulate matter deposition in the drainage pipe network, and preprocesses the data to ensure that the data quality meets the accuracy requirements for predicting particulate matter deposition in the drainage pipe network. A step prediction framework for particulate matter deposition in the drainage pipe network is constructed, and the framework contains m models for realizing m step prediction of particulate matter deposition in the drainage pipe network. The network unit parameters are selected and the prediction performance evaluation indexes for the particulate matter deposition prediction model in the drainage pipe network are determined, and the step prediction framework for particulate matter deposition in the drainage pipe network is trained to ensure that the framework can accurately simulate the change process of particulate matter deposition in the drainage pipe network. A double-layer prediction framework for particulate matter deposition in the drainage pipe network is constructed and compared with the prediction performance of the step prediction framework for particulate matter deposition in the drainage pipe network.

Description

Technical Field

[0001] The present invention belongs to the technical field of predicting particulate matter deposition in drainage pipe networks of drainage pipe networks, and specifically relates to a method for predicting particulate matter deposition in drainage pipe networks based on a long short-term memory step sequence. Background Art

[0002] The deposition of particulate matter in the drainage pipe network will affect the hydraulic capacity of the urban drainage and sewer system, and affect the operation efficiency of the drainage pipe network. In order to reduce the adverse effects brought by the continuous deposition of particulate matter in the drainage pipe network, the drainage pipe network system is usually designed with a self-cleaning mechanism to keep the bottom of the drainage pipe network clean and prevent particulate matter from depositing in the drainage pipe network. Obviously, accurate prediction and simulation of particulate matter deposition in the drainage pipe network can better prevent the deposition of particulate matter and maintain the normal operation of the drainage pipe network.

[0003] The deposition of particulate matter in the drainage pipe network affects both the flow velocity in the drainage pipe network and the shear stress distribution at the pipe wall. The main reasons for particulate matter deposition include water flow irregularity, industrial emissions, and infiltration. If the sediment stays in the drainage pipe network for too long, the properties of the sediment may change, leading to further deposition and causing the complete blockage of the drainage pipe network. Therefore, accurately predicting the deposition of particulate matter in the drainage pipe network and taking effective countermeasures are of great significance for improving the operation efficiency of the drainage pipe network.

[0004] Drainage pipe networks can usually be regarded as rigid boundary channels, which are channels with an immovable bed and side walls. Currently, in order to predict the deposition of particulate matter in the drainage pipe network, the Froude number of particulate matter in the drainage pipe network is usually predicted, and the Froude number is used to reflect the deposition of particulate matter. In recent years, data-driven methods such as artificial neural networks, recurrent neural networks, and long short-term memory networks (LSTM) have been widely used in urban water systems. These methods can use big data to simulate highly nonlinear and complex systems and have good performance in the hydrological field.

[0005] Artificial neural networks are widely used to process the modeling of complex systems in the hydrological field, such as evapotranspiration prediction, flood prediction, and rainfall-runoff modeling. However, many traditional artificial neural network-based methods do not consider the time series information of input data. Compared with traditional artificial neural networks, recurrent neural networks fully consider the time series information of input data, can remember the information of input data before, and capture the time dynamic change information, so as to obtain better prediction performance. Obviously, compared with traditional artificial neural networks, recurrent neural networks are more suitable for modeling complex time series such as hydrology and meteorology.

[0006] However, when traditional recurrent neural networks are trained in the presence of long lags, problems of exploding and vanishing gradients occur. The training of traditional recurrent neural networks relies on an extended version of backpropagation called backpropagation through time, which requires not only computing the cost gradient corresponding to the input weights but also computing the hidden weights corresponding to previous time steps. The cost gradient can explode or vanish with long lags, which results in the poor update of the learnable parameters of traditional recurrent neural networks. Long short-term memory networks (LSTMs) can overcome the above drawbacks because they maintain their state over time, and their non-linear gating units (i.e., forget gate, input gate, and output gate) regulate the flow of information into and out of the storage unit. The forget gate determines what information the unit state will forget. Its range is from 0 to 1, which means "completely delete memory" to "completely retain memory". The input gate determines what information is used to update the unit state. The output gate controls the unit state information flowing into the new hidden state. In addition, in recent years, deep learning has received extensive attention, and deep neural networks have achieved excellent performance in difficult tasks in different fields. The original long short-term memory network and its variants (e.g., gated recurrent unit), which are types of deep neural networks configured with a recurrent neural network framework, have shown remarkable performance in language modeling, speech-to-text transcription, machine translation, and other applications.

[0007] In the hydrological field, methods based on long short-term memory networks have shown good performance. For example, long short-term memory networks are used to simulate and predict the water level of combined sewer overflows and predict the groundwater level in areas where hydrogeological data are difficult to obtain. However, traditional long short-term memory network-based methods do not encode the in-pipe particulate matter in the drainage pipe network as input data and do not obtain the following important information in the input data: (a) the sequential information of the in-pipe particulate matter sequence in the drainage pipe network; (b) the previous particulate matter deposition situation containing partial compressed information of all factors affecting the predicted particulate matter deposition. The compressed information mainly includes all possible influencing factors affecting runoff pollution, such as the volume sand content under non-deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment diameter ( ), hydraulic radius ( ), and pipe friction coefficient ( ), etc. The term "compressed" means the low-dimensional results obtained from high-dimensional factors, which contain the information of these factors. The sequence-to-sequence model does not consider the following situation in practice: the particulate matter deposition on the th day is not affected by the operation status of the drainage pipe network after the th day. Summary of the Invention

[0008] To solve the current existing technical problems, the present invention aims to propose a method for predicting particulate matter deposition in the drainage pipe network based on a long short-term memory step sequence, which is used for predicting particulate matter deposition in the drainage pipe network. The long short-term memory network solves the problems of gradient explosion and disappearance in predicting particulate matter deposition in the drainage pipe network, can improve the prediction performance, and improving the prediction accuracy of particulate matter deposition in the drainage pipe network is of great significance for improving the operation efficiency of the drainage pipe network. At the same time, a step prediction framework for particulate matter deposition sequence in the drainage pipe network, which framework contains m models, and can achieve m step prediction of particulate matter deposition in the drainage pipe network. When predicting particulate matter deposition in the drainage pipe network, the current in-pipe particulate matter deposition situation is not affected by the later in-pipe conditions, which is more in line with the actual situation; moreover, the Froude number of particulate matter in the drainage pipe network can reflect the deposition situation of particulate matter, and is also one of the key indicators concerned by the self-cleaning mechanism of the drainage pipe network.

[0009] To achieve the above technical features, the object of the present invention is achieved as follows: A method for predicting particulate matter deposition in the drainage pipe network based on a long short-term memory step sequence, comprising the following steps:

[0010] Step 1, prepare the input data and output data for predicting particulate matter deposition in the drainage pipe network, and preprocess the data;

[0011] Step 2, construct a prediction model for particulate matter deposition in the drainage pipe network based on a long short-term memory step sequence, and predict particulate matter deposition in the drainage pipe network;

[0012] Step 3, train and test the model, set the model parameters to ensure that the model can accurately simulate particulate matter deposition in the drainage pipe network;

[0013] Step 4, simulate particulate matter deposition in the drainage pipe network, and evaluate the prediction performance of the model.

[0014] The specific steps of the said Step 1 are as follows:

[0015] Step 1.1, select the main influencing factors for predicting particulate matter deposition in the drainage pipe network, that is, the input data, and the decision variable, that is, the output data; wherein, the influencing factors mainly include: the volume sand content under the non-depositing condition of the deposition bed , dimensionless particle diameter , median sediment particle size , hydraulic radius and pipe friction coefficient ; the output data of the prediction model is the Froude number of the particulate matter ;

[0016] Step 1.2, after collecting all the input data and output data of the prediction model for particulate matter deposition in the drainage pipe network, check the data quality, and use a sliding time window to replace the missing data and incorrect data in the input data and output data to ensure that the quality of the input data and output data of the prediction model meets the prediction requirements.

[0017] The specific steps of the said Step 2 are as follows:

[0018] Step 2.1, construct a step prediction framework for particulate matter deposition in the drainage pipe network based on a long short-term memory network, and this framework contains m models for realizing m step prediction of particulate matter deposition in the drainage pipe network;

[0019] Step 2.2, use the first model in the step prediction framework for particulate matter deposition in the drainage pipe network to predict the particulate matter deposition in the drainage pipe network one step ahead, and use it as the input of the prediction models for particulate matter deposition in the drainage pipe network two steps, three steps, …, m steps ahead, that is, the second, third, …, m models in the step prediction framework for particulate matter deposition in the drainage pipe network;

[0020] Step 2.3, use the i (1 < i < m )th model in the step prediction framework for particulate matter deposition in the drainage pipe network to predict the particulate matter deposition in the drainage pipe network i steps ahead, and use it as the input of the prediction models for particulate matter deposition in the drainage pipe network i + 1, …, m steps ahead, that is, the i + 1th, …, m models in the step prediction framework for particulate matter deposition in the drainage pipe network;

[0021] Step 2.4, use the m th model in the step prediction framework for particulate matter deposition in the drainage pipe network to predict the particulate matter deposition in the drainage pipe network m steps ahead;

[0022] Step 2.5, combining Steps 2.2 - 2.4, predict the particulate matter deposition prediction data in the drainage pipe network m steps ahead through m models.

[0023] The specific steps of the said Step 3 are as follows:

[0024] Step 3.1, determine the learnable parameters and hyperparameters of the particulate matter deposition prediction model in the drainage pipe network to train and test the m models in the step prediction framework of the particulate matter sequence in the drainage pipe network, and ensure that the prediction model can accurately output the particulate matter deposition prediction data in the drainage pipe network;

[0025] Step 3.2, determine the performance evaluation indicators of the particulate matter deposition prediction model in the drainage pipe network to effectively train each model in the step prediction framework of the particulate matter deposition in the drainage pipe network and obtain a particulate matter deposition prediction model in the drainage pipe network that meets the accuracy requirements;

[0026] Step 3.3, input the input data of the particulate matter deposition prediction model in the drainage pipe network into the particulate matter deposition prediction model framework in the drainage pipe network, train and test the particulate matter deposition prediction model in the drainage pipe network, and ensure that the output data of the particulate matter deposition prediction model in the drainage pipe network can accurately reflect the actual particulate matter deposition change in the drainage pipe network.

[0027] The specific content of the said Step 4 is as follows:

[0028] Step 4.1, construct a two-layer long short-term memory network as the particulate matter deposition prediction benchmark model in the drainage pipe network to evaluate the prediction performance of the step prediction framework of the particulate matter deposition sequence in the drainage pipe network;

[0029] Step 4.2, predict the particulate matter deposition time series data in the drainage pipe network through the volume sand content , dimensionless particle diameter , sediment median diameter , hydraulic radius and pipe friction coefficient under the non-deposition condition of the deposition bed in the input data of the particulate matter deposition prediction in the drainage pipe network, and compare it with the output result of the step prediction framework of the particulate matter deposition sequence in the drainage pipe network;

[0030] Step 4.3, through the input data of the particulate matter deposition prediction model in the drainage pipe network: volume sand content , dimensionless particle diameter , sediment median diameter , hydraulic radius and pipe friction coefficient , predict the output data of the particulate matter deposition prediction model in the drainage pipe network: Froude number of the particulate matter, and compare it with the Froude number calculated by the empirical formula to analyze the prediction performance of the model.

[0031] The specific content of the said Step 2.1 is as follows:

[0032] Step 2.1.1, construct a step prediction model for particulate matter deposition in a single drainage pipe network. Predict the output data of particulate matter deposition in the drainage pipe network through the input data of particulate matter deposition prediction in the drainage pipe network;

[0033] Step 2.1.2, construct step prediction models for particulate matter deposition in the drainage pipe network to obtain a step prediction framework for particulate matter deposition in the drainage pipe network, including step prediction models for particulate matter sequences in the drainage pipe network. Use the input data of particulate matter deposition prediction in the drainage pipe network to predict and obtain steps of particulate matter deposition prediction data in the drainage pipe network;

[0034] The specific content of step 2.5 is as follows:

[0035] Step 2.5.1, train the first lead particulate matter deposition prediction model in the drainage pipe network. Predict the output data of the first-step lead particulate matter deposition in the drainage pipe network through the input data of particulate matter deposition prediction in the drainage pipe network;

[0036] Step 2.5.2, train the second lead particulate matter deposition prediction model in the drainage pipe network. Predict the output data of the second-step lead particulate matter deposition in the drainage pipe network through the input data of particulate matter deposition prediction in the drainage pipe network and the output data of the first-step lead particulate matter deposition in the drainage pipe network;

[0037] Step 2.5.3, train the i ( ) lead particulate matter deposition prediction model in the drainage pipe network. Predict the output data of the i -step lead particulate matter deposition in the drainage pipe network through the input data of particulate matter deposition prediction in the drainage pipe network and the output data of the previous i -step lead particulate matter deposition prediction in the drainage pipe network;

[0038] Step 2.5.4, summarize the output data of the m-step particulate matter deposition prediction in the drainage pipe network obtained in steps 2.5.1 - 2.5.3 to obtain the m-step lead particulate matter deposition prediction data in the drainage pipe network.

[0039] The specific content of step 3.1 is as follows:

[0040] Step 3.1.1, determine the dimensions of each long short-term memory network unit. Each network unit contains 5 dimensions, including: the volume sand content under non-deposition conditions of the sediment bed , dimensionless particle diameter , median sediment particle size , hydraulic radius and the pipe friction coefficient ;

[0041] Step 3.1.2, normalize the input data of 5 dimensions for each long short-term memory network unit, and all input data are normalized by the z-score normalization method;

[0042] Step 3.1.3, set the hidden state length of each long short-term memory network unit to 32, set the dropout rate to 0.1 to avoid overfitting, set the learning rate to 0.001, set the number of samples per batch to 256, and use the root mean square error MSE as the loss function;

[0043] Step 3.1.4, select the true observation data of the in-pipe particulate matter deposition in the drainage pipe network for k times to train the proposed step prediction framework for the in-pipe particulate matter deposition sequence in the drainage pipe network. Through k times of training, k NSE values are obtained, and select the hyperparameter set that produces the highest NSE median among all hyperparameter combinations as the hyperparameters of the step prediction framework for the in-pipe particulate matter deposition sequence in the drainage pipe network.

[0044] The specific content of step 4.1 is as follows:

[0045] Step 4.1.1, use a long short-term memory network to construct a two-layer prediction framework for the in-pipe particulate matter deposition sequence in the drainage pipe network. This framework includes two encoder sequences, a state vector, and a decoder sequence. The first decoder sequence has long short-term memory network units, and its input includes the volume sand content under the non-deposition condition of the sediment bed , dimensionless particle diameter , median sediment particle size , hydraulic radius and the pipe friction coefficient ; The second decoder sequence has long short-term memory network units, and its input data is the observed data of the in-pipe particulate matter deposition in the drainage pipe network; the state vector contains the compressed and processed information from the inputs of these two encoder sequences; the decoder sequence has m long short-term memory network units, and the output is the m-step-ahead prediction of the in-pipe particulate matter deposition in the drainage pipe network;

[0046] Step 4.1.2, use the actual data of the in-pipe particulate matter deposition in the drainage pipe networks of k cities to train the two-layer prediction framework for the in-pipe particulate matter deposition sequence in the drainage pipe network. Through k times of training, k NSE values are obtained, and select the hyperparameter set that produces the highest NSE median among all hyperparameter combinations as the hyperparameters of the two-layer prediction framework for the in-pipe particulate matter deposition sequence in the drainage pipe network.

[0047] The specific content of step 4.2 is as follows:

[0048] Step 4.2.1, select The actual observed data of particulate matter deposition in the drainage pipe network of a city are respectively input into the step prediction framework of particulate matter deposition sequence in the drainage pipe network and the prediction framework of particulate matter deposition sequence in the double-layer drainage pipe network to obtain the predicted data of particulate matter deposition in the drainage pipe network;

[0049] Step 4.2.2, after obtaining the predicted data of particulate matter deposition in the drainage pipe network, according to the Froude number output by the particulate matter deposition prediction model in the drainage pipe network, calculate the consistency index, mean absolute error, root mean square error, coefficient of determination and adjustment coefficient in sequence, and compare and evaluate the prediction performance of the step prediction framework of particulate matter deposition sequence in the drainage pipe network and the prediction framework of particulate matter deposition sequence in the double-layer drainage pipe network based on these indicators.

[0050] The present invention has the following beneficial effects:

[0051] The prediction method of particulate matter deposition in the drainage pipe network based on the long short-term memory step sequence is of great significance for improving the prediction accuracy of particulate matter deposition in the drainage pipe network, especially for realizing multi-step advanced prediction of particulate matter deposition in the drainage pipe network. The long short-term memory network solves the problems of gradient explosion and disappearance in the prediction of particulate matter deposition in the drainage pipe network and can improve the prediction performance. At the same time, the step prediction framework of particulate matter deposition sequence in the drainage pipe network, which framework contains m models, can realize m step prediction of particulate matter deposition in the drainage pipe network. When predicting the particulate matter deposition in the drainage pipe network, the current particulate matter deposition in the pipe is not affected by the later pipe conditions, which is more in line with the actual situation. Description of the Drawings

[0052] The present invention will be further described below with reference to the drawings and embodiments.

[0053] Figure 1 It is a flow chart of the prediction method of particulate matter deposition in the drainage pipe network based on the long short-term memory step sequence.

[0054] Figure 1 It is mainly divided into 4 parts: 1. Preparation of predicted data of particulate matter deposition in the drainage pipe network; 2. Construction of the step prediction framework of particulate matter deposition sequence in the drainage pipe network; 3. Step prediction framework of particulate matter deposition sequence in the drainage pipe network; 4. Comparison and evaluation of the prediction performance of particulate matter prediction in the drainage pipe network.

[0055] Figure 2 It is a schematic diagram of the network structure of the recurrent neural network.

[0056] Figure 3 It is a schematic diagram of the network structure of the long short-term memory network.

[0057] Figure 4 Schematic diagram of the step prediction framework for particulate matter deposition in the drainage pipe network

[0058] Figure 5 Schematic diagram of the system structure of the prediction framework for the double-layer rainfall runoff pollution sequence

[0059] Figure 6 Schematic diagram of the network structure of the prediction framework for the double-layer rainfall runoff pollution sequence

[0060] Figure 7 Frame diagram of the prediction method for particulate matter deposition in the drainage pipe network based on the long short-term memory step sequence

[0061] Figure 8 Prediction effect diagram when using the first group of non-deposition data to train the step prediction framework for particulate matter deposition in the drainage pipe network

[0062] Figure 9 Prediction effect diagram when using the second group of non-deposition data to train the step prediction framework for particulate matter deposition in the drainage pipe network

[0063] Figure 10 Prediction effect diagram when using the third group of non-deposition data to train the step prediction framework for particulate matter deposition in the drainage pipe network

[0064] Figure 11 Prediction effect diagram when using the fourth group of non-deposition data to train the step prediction framework for particulate matter deposition in the drainage pipe network

[0065] Figure 12 Prediction effect diagram when using the fifth group of non-deposition data to train the step prediction framework for particulate matter deposition in the drainage pipe network

[0066] Figure 13 Model prediction performance obtained by the prediction model, empirical model, and two empirical formulas using the first group of real observation data

[0067] Figure 14 Model prediction performance obtained by the prediction model, empirical model, and two empirical formulas using the second group of real observation data

[0068] Figure 15 Model prediction performance obtained by the prediction model, empirical model, and two empirical formulas using the third group of real observation data

[0069] Figure 16 Model prediction performance obtained by the prediction model, empirical model, and two empirical formulas using the fourth group of real observation data

[0070] Figure 17The model prediction performances obtained by the prediction model, empirical model, and two empirical formulas using the fifth set of real observation data. Detailed implementation manners

[0071] The following further describes the implementation manners of the present invention with reference to the accompanying drawings.

[0072] Example 1:

[0073] Refer to Figure 1 , the present invention provides a prediction method for particulate matter deposition in a drainage pipe network based on a long short-term memory step sequence, including the following steps:

[0074] Step 1, prepare the input data and output data for predicting particulate matter deposition in the drainage pipe network, and preprocess the data;

[0075] Step 2, construct a prediction model for particulate matter deposition in the drainage pipe network based on a long short-term memory step sequence, and predict the deposition process of particulate matter in the drainage pipe network;

[0076] Step 3, train and test the model, and set the model parameters to ensure that the model prediction results can accurately reflect the deposition process of particulate matter in the drainage pipe network;

[0077] Step 4, predict the deposition process of particulate matter in the drainage pipe network, and evaluate the prediction performance of the model;

[0078] Furthermore, the specific content of step 1 is as follows:

[0079] Step 1.1, select the main influencing factors (input quantities) and decision variables (output quantities) for predicting particulate matter deposition in the drainage pipe network. The influencing factors mainly include: the volume sand content under non-deposition conditions of the deposition bed ( ), the dimensionless particle diameter ( ), the median particle size of the sediment ( ), the hydraulic radius ( ), and the pipe friction coefficient ( ). The output quantity of the prediction model is the Froude number of the particulate matter ( ).

[0080] Step 1.2, after collecting all the input data and output data of the prediction model for particulate matter deposition in the drainage pipe network, check the data quality, and use a sliding time window to replace the missing data and incorrect data in the input data and output data to ensure that the quality of the input data and output data of the prediction model meets the prediction requirements;

[0081] Furthermore, the specific content of step 2 is as follows:

[0082] Step 2.1, construct a step prediction framework for the particulate matter deposition sequence in the drainage pipe network based on a long short-term memory network. This framework contains m models for realizing m step prediction of particulate matter deposition in the drainage pipe network;

[0083] Step 2.2, use the first model in the step prediction framework for the particulate matter deposition sequence in the drainage pipe network to predict the predicted value of the particulate matter deposition in the drainage pipe network with the first-step lead, and use it as the input of the predicted models for the particulate matter deposition in the drainage pipe network with the (second, third,..., m )-step lead (i.e., the second, third,..., m models in the particulate matter deposition prediction framework for the drainage pipe network);

[0084] Step 2.3, use the i (1 < i < m )-th model in the step prediction framework for the particulate matter deposition sequence in the drainage pipe network to predict the predicted value of the particulate matter deposition in the drainage pipe network with the i -step lead, and use it as the input of the predicted models for the particulate matter deposition in the drainage pipe network with the ( i + 1,..., m )-step lead (i.e., the i + 1-th,..., m -th models in the particulate matter deposition prediction framework for the drainage pipe network);

[0085] Step 2.4, use the m -th model in the particulate matter deposition prediction framework for the drainage pipe network to predict the predicted value of the particulate matter deposition in the drainage pipe network with the m -step lead;

[0086] Step 2.5, combining Steps 2.2 - 2.4, through m models, predict the predicted data of the particulate matter deposition in the drainage pipe network with the m -step lead.

[0087] Furthermore, the specific content of Step 3 is as follows:

[0088] Step 3.1, determine the learnable parameters and hyperparameters of the particulate matter deposition prediction model in the drainage pipe network, so as to train and test the m models in the step prediction framework for the particulate matter deposition sequence in the drainage pipe network, and ensure that the prediction model can accurately output the particulate matter deposition sequence data in the drainage pipe network;

[0089] Step 3.2, determine the performance evaluation index of the particulate matter deposition prediction model in the drainage pipe network to ensure that each model in the step prediction framework of particulate matter deposition in the drainage pipe network can be effectively trained, and a particulate matter deposition prediction model in the drainage pipe network that meets the accuracy requirements can be obtained;

[0090] Step 3.3, input the input data of the particulate matter deposition prediction model in the drainage pipe network into the step prediction framework of particulate matter deposition in the drainage pipe network, and train and test the particulate matter deposition prediction model in the drainage pipe network to ensure that the output data of the particulate matter deposition prediction model in the drainage pipe network (i.e., the Froude number of the particulate matter in the drainage pipe network) can accurately reflect the actual particulate matter deposition situation in the drainage pipe network.

[0091] Furthermore, the specific content of step 4 is as follows:

[0092] Step 4.1, construct a two-layer long short-term memory network as the particulate matter deposition prediction benchmark model in the drainage pipe network to evaluate the prediction performance of the step prediction framework of particulate matter deposition in the drainage pipe network;

[0093] Step 4.2, through the input data of the particulate matter deposition prediction model in the drainage pipe network (volume sand content under non-deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( )) to predict the output data of the particulate matter deposition prediction model in the drainage pipe network (Froude number of the particulate matter ( )) and compare it with the output result of the step prediction framework of particulate matter deposition in the drainage pipe network;

[0094] Step 4.3, through the input data of the particulate matter deposition prediction model in the drainage pipe network (volume sand content under non-deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( )) to predict the output data of the particulate matter deposition prediction model in the drainage pipe network (Froude number of the particulate matter ( )) and compare it with the Froude number calculated by the empirical formula to analyze the prediction performance of the model.

[0095] When making the comparison, the following two empirical formulas are mainly considered:

[0096] ;

[0097] ;

[0098] Furthermore, step 2.1 is specifically as follows;

[0099] Step 2.1.1: Construct a single-step prediction model for particulate matter deposition in the drainage pipe network. Predict the particulate matter deposition value in the drainage pipe network through the input data of particulate matter deposition prediction in the drainage pipe network.

[0100] Long short-term memory network belongs to the recurrent neural network. It takes the state of the hidden neurons in the previous time step as an additional input in the next time step, remembers the previous information, and captures the time dynamic changes. In the traditional recurrent neural network unit, the state of the output data can be expressed as:

[0101] ;

[0102] In the formula, is the hyperbolic tangent function, is the input vector at time is the hidden state at time is the learnable matrix connecting two hidden states, is the learning matrix connecting the input vector and the hidden state, is the adjustable bias vector. Due to its gradient vanishing or exploding problem, this model cannot learn the patterns with long-term lags well;

[0103] The long short-term memory network can overcome this shortcoming by replacing the ordinary neurons of the recurrent neural network with a storage block containing four parts: in addition to three gates (i.e., the forget gate, the input gate, and the output gate), it also includes a constant error carousel unit. The constant error carousel unit can run straight down along the entire chain without any activation function. Therefore, when training the long short-term memory network using the time reverse propagation of particulate matter deposition sequence in the drainage pipe network, the gradient will not vanish. In addition, the long short-term memory network can maintain its state over time, and its non-linear gating unit can regulate the information flow into and out of the storage unit;

[0104] The working principle of the long short-term memory network can be expressed by the following mathematical equation:

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] wherein, is the sigmoid function, represents element-wise multiplication, is the input gate at time , and its learnable parameter is , is the forget gate at time , and its learnable parameter is , is the output gate at time , and its learnable parameter is , is the latent state of the cell at time , and its learnable parameter is , is the cell state at time , is the hidden state of the cell. These gates can control the flow of information into and out of the storage cell. The forget gate determines what information the cell state will forget. The value range is from 0 to 1, indicating "completely delete memory" to "completely retain memory". The input gate determines which information to use to update the cell state. The latent cell state gate (i.e., the tanh layer) adds a new candidate value to the state. The output gate controls the flow of information from the cell state into the new hidden state;

[0112] Step 2.1.2 Construct step prediction models for particulate matter deposition in the drainage pipe network to obtain a step prediction framework for particulate matter deposition in the drainage pipe network (including step prediction models for particulate matter deposition in the drainage pipe network), and use the input data for particulate matter deposition prediction in the drainage pipe network to predict steps of particulate matter deposition prediction data in the drainage pipe network;

[0113] The step prediction framework for particulate matter deposition in the drainage pipe network includes step prediction models for particulate matter deposition in the drainage pipe network, which are used for step particulate matter deposition prediction in the drainage pipe network. The step prediction models for particulate matter deposition in the drainage pipe network are not independent of each other. For example, when training the prediction model for particulate matter deposition in the drainage pipe network at the second step, the output data of the prediction model for particulate matter deposition in the drainage pipe network at the first step is required; for the prediction model for particulate matter deposition in the drainage pipe network at the previous step, the input data is composed of the last consists of consecutive time steps, where ( ) includes the volume sand content under non-deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ) and pipe friction coefficient ( );

[0114] ;

[0115] In the formula, is the predicted value of particulate matter deposition in the drainage pipe network on the th day. Using the observed data of particulate matter deposition in the drainage pipe network and days of predicted input data for particulate matter deposition in the drainage pipe network (volume sand content under non-deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ) and pipe friction coefficient ( )) to predict the particulate matter deposition in the drainage pipe network on the th day (i.e., the particulate matter deposition in the drainage pipe network in the first step). Then, the prediction model for particulate matter deposition in the second-step advanced drainage pipe network is defined as:

[0116] ;

[0117] Using times of observed data of particulate matter deposition in the drainage pipe network, times of prediction of particulate matter deposition in the drainage pipe network and days of data on influencing factors of particulate matter deposition in the drainage pipe network, the particulate matter deposition in the drainage pipe network on the th day (i.e., the particulate matter deposition in the drainage pipe network in the second step) is predicted. Similarly, the prediction model for particulate matter deposition in the th-step advanced drainage pipe network is defined as:

[0118] ;

[0119] Using days of observed data of particulate matter deposition in the drainage pipe network, days of predicted data of particulate matter deposition in the drainage pipe network and days of observed data on influencing factors of particulate matter deposition in the drainage pipe network, the particulate matter deposition in the drainage pipe network on the th day (i.e., the particulate matter deposition in the drainage pipe network in the rd step) is predicted.

[0120] Furthermore, the specific content of step 2.5 is as follows:

[0121] Step 2.5.1: Train the first prediction model for particulate matter deposition in the advanced drainage pipe network. Use the volume sand content under non-depositing conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( ) to predict the particulate matter prediction data in the first-step advanced drainage pipe network;

[0122] Step 2.5.2: Train the second prediction model for particulate matter deposition in the advanced drainage pipe network. Use the volume sand content under non-depositing conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( ) and the particulate matter deposition prediction data in the first-step advanced drainage pipe network;

[0123] Step 2.5.3: Train the i ( )-th prediction model for particulate matter deposition in the advanced drainage pipe network. Use the volume sand content under non-depositing conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( ) and the particulate matter prediction data in the previous i - 1-step drainage pipe network to predict the particulate matter deposition prediction output data in the -th step of the advanced drainage pipe network;

[0124] Step 2.5.4: Aggregate the particulate matter deposition prediction output data in steps 2.5.1 - 2.5.3 to obtain the particulate matter deposition prediction data in the m -th step of the advanced drainage pipe network. m

[0125] Furthermore, the specific content of step 3.1 is as follows:

[0126] Step 3.1.1: Determine the dimensions of each long short-term memory network unit. Each network unit contains 4 dimensions, including: the volume sand content under non-depositing conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( );

[0127] Step 3.1.2: Normalize the input data of the 4 dimensions of each long short-term memory network unit. All input data is normalized using the z-score normalization method;

[0128] Step 3.1.3: Set the hidden state length of each long short-term memory network unit to 32. To avoid overfitting, set the dropout rate to 0.1, the learning rate to 0.001, the batchsize (number of samples per batch) to 256, and use the mean squared error (MSE) as the loss function;

[0129] Step 3.1.4: Select k true observed data of particulate matter deposition in the drainage pipe network for multiple times to train the proposed step prediction framework for particulate matter deposition sequence in the drainage pipe network. Through k multiple trainings, obtain k NSE values. Select the set of hyperparameters that produce the highest median NSE among all hyperparameter combinations as the hyperparameters of the step prediction framework for particulate matter deposition sequence in the drainage pipe network.

[0130] Furthermore, the specific content of Step 4.1 is as follows:

[0131] Step 4.1.1: Use a long short-term memory network to construct a two-layer prediction framework for particulate matter deposition sequence in the drainage pipe network. This framework includes two encoder sequences, a state vector, and a decoder sequence. The first decoder sequence has long short-term memory network units, and its inputs include the volume sand content under non-deposition conditions of the deposition bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( ). The second decoder sequence has long short-term memory network units, and its input data is the observed data of particulate matter deposition in the drainage pipe network. The state vector contains the compressed and processed information from the inputs of these two encoder sequences. The decoder sequence has m long short-term memory network units, and the output is m step-ahead prediction of particulate matter deposition in the drainage pipe network;

[0132] Step 4.1.2: Use k actual data of particulate matter deposition prediction in the drainage pipe network to train the two-layer prediction framework for particulate matter deposition sequence in the drainage pipe network. Through k multiple trainings, obtaink For the NSE value, select the set of hyperparameters that produces the highest median NSE among all combinations of hyperparameters as the hyperparameters for the prediction framework of the in-pipe particulate matter deposition sequence in the double-layer drainage pipe network.

[0133] Furthermore, step 4.2 is specifically as follows:

[0134] In step 4.2.1, select the actual observed data of the in-pipe particulate matter deposition in the drainage pipe networks of several cities, and input them into the step prediction framework of the in-pipe particulate matter deposition sequence in the drainage pipe network and the prediction framework of the in-pipe particulate matter deposition sequence in the double-layer drainage pipe network respectively to obtain the predicted data of the in-pipe particulate matter deposition in the drainage pipe network;

[0135] In step 4.2.2, after obtaining the predicted data of the in-pipe particulate matter deposition in the drainage pipe network, calculate the consistency index, mean absolute error, root mean square error, coefficient of determination and adjustment coefficient in sequence according to the Froude number output by the in-pipe particulate matter deposition prediction model in the drainage pipe network, and compare and evaluate the prediction performance of the step prediction framework of the in-pipe particulate matter deposition sequence in the drainage pipe network and the prediction framework of the in-pipe particulate matter deposition sequence in the double-layer drainage pipe network based on these indicators;

[0136] The calculation formula for the consistency index of the Froude number is:

[0137] ;

[0138] In the formula, represents the consistency index of the Froude number, which is the ratio of the root mean square error to the potential error and is used to measure the quality of the model. The value of varies between 0 and 1, where 1 represents a perfect match and 0 represents a complete mismatch. represents the actual value of the Froude number, represents the predicted value of the Froude number,

[0139] The calculation formula for the mean absolute error of the Froude number is:

[0140] ;

[0141] In the formula, MAE represents the mean absolute error of the Froude number, which is the average of the absolute errors between all predicted values and true values of the Froude number. The value range of MAE is from 0 to , and the smaller the better;

[0142] The calculation formula for the root mean square error of the Froude number is:

[0143] ;

[0144] In the formula, represents the root mean square error of the Froude number;

[0145] For the formula for calculating the coefficient of determination is:

[0146] ;

[0147] determines the proportion of the variance in the actual values that can be explained by the predicted variable. It can be estimated as a combined distribution of the single distributions of the actual and predicted values. It is a measure of the goodness of fit of a statistical model. The value ranges from 0 to 1, where 1 represents a good model and 0 represents a biased model;

[0148] For the formula for calculating the adjustment coefficient is:

[0149] Adjusted ;

[0150] Example 2:

[0151] The object of the present invention is to propose a prediction method for particulate matter deposition in a drainage pipe network based on a long short-term memory step sequence. The long short-term memory network solves the problems of gradient explosion and disappearance in the prediction of particulate matter deposition in a drainage pipe network and can improve the prediction performance. At the same time, a step prediction framework for particulate matter deposition in a drainage pipe network, which framework contains m models and can achieve m step prediction of particulate matter deposition in a drainage pipe network. When predicting the particulate matter deposition in a drainage pipe network, the current particulate matter deposition situation in the pipe is not affected by the later pipe conditions and is more in line with the actual situation.

[0152] Five groups of non-deposited layer data commonly used in existing research are selected to illustrate the proposed method. The input data includes the volume sand content under non-deposited conditions of the sediment bed ( ), dimensionless particle diameter ( ), median particle size of the sediment ( ), hydraulic radius ( ), and pipe friction coefficient ( ), and the output of the prediction model is the Froude number of the particulate matter ( );

[0153] The main influencing factors of the particulate matter deposition prediction model in the drainage pipe network are shown in Table 1:

[0154] Table 1 Main influencing factors of the particulate matter deposition prediction model in the drainage pipe network

[0155]

[0156] The data ranges of the 5 groups of non-deposited layer data are shown in Table 2:

[0157] Table 2 Data ranges of the 5 groups of non-deposited layer data

[0158]

[0159] After collecting the 5 groups of non-deposited layer data, the sliding time window is used to replace the missing data and incorrect data in the volume sand content ( ), dimensionless particle diameter ( ), sediment median grain size ( ), hydraulic radius ( ), pipe friction coefficient ( ), and Froude number ( ) under the non-deposited condition of the sediment bed in turn, so as to ensure that the quality of the input data and output data of the particulate matter deposition prediction model in the drainage pipe network meets the prediction requirements;

[0160] As Figure 2 shown, the long short-term memory network belongs to the recurrent neural network, which takes the state of the hidden neurons in the previous time step as an additional input in the next time step, remembers the previous information and captures the time dynamic changes. denotes the input at time , denotes the hidden state at time . In the traditional recurrent neural network unit, the state of the output data can be expressed as:

[0161] ;

[0162] In the formula, is the hyperbolic tangent function, is the input vector at time is the hidden state at time is the learnable matrix connecting the two hidden states, is the learning matrix connecting the input vector and the hidden state, is the adjustable bias vector. Due to its gradient vanishing or explosion problem, this model cannot learn the patterns with long-term lags well;

[0163] The long short-term memory network can overcome this shortcoming by replacing the ordinary neurons of the recurrent neural network with a memory block consisting of four parts: in addition to the three gates (i.e., the forget gate, the input gate, and the output gate), it also includes a constant error carousel unit. The constant error carousel unit can run straight down the entire chain without any activation function, so the gradient will not disappear when training the long short-term memory network using the time reverse propagation of the particulate matter deposition sequence in the drainage pipe network. In addition, the long short-term memory network can maintain its state over time, and its non-linear gating unit can regulate the information flow into and out of the memory unit;

[0164] As Figure 3 shown, 、 and respectively represent the input, the hidden state, and the cell state at time

[0165] The working principle of the long short-term memory network can be expressed by the following mathematical equations:

[0166] ;

[0167] ;

[0168] ;

[0169] ;

[0170] ;

[0171] ;

[0172] In the formula, is the sigmoid function, represents element-wise multiplication, is the input gate at time , and its learnable parameter is , is the forget gate at time , and its learnable parameter is , is the output gate at time , and its learnable parameter is , is the potential state of the cell at time , and its learnable parameter is , is the cell state at time , is the hidden state of the cell. These gates can control the flow of information into and out of the memory cell. The forget gate determines what information the cell state will forget. It ranges from 0 to 1, representing "completely delete memory" to "completely retain memory". The input gate determines which information to use to update the cell state. The potential cell state gate (i.e., the tanh layer) adds a new candidate value to the state. The output gate controls the flow of information from the cell state into the new hidden state;

[0173] Step 2.1.2, construct step prediction models for particulate matter deposition in the drainage pipe network, and obtain the step prediction framework for particulate matter deposition in the drainage pipe network (including step prediction models for particulate matter deposition in the drainage pipe network). Use the input data for particulate matter deposition prediction in the drainage pipe network to predict step prediction data for particulate matter deposition in the drainage pipe network, as Figure 4 shown;

[0174] The step prediction framework for particulate matter deposition in the drainage pipe network includes step prediction models for particulate matter deposition in the drainage pipe network, which are used for step prediction of particulate matter deposition in the drainage pipe network. The step prediction models for particulate matter deposition in the drainage pipe network are not independent of each other. For example, when training the step-ahead prediction model for particulate matter deposition in the drainage pipe network at the second step, the output data of the step-ahead prediction model for particulate matter deposition in the drainage pipe network at the first step are required; for the previous step prediction model for particulate matter deposition in the drainage pipe network, the input data consists of the last consecutive time steps of the variable, where ( ) includes the volume sand content under non-deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ) and pipe friction coefficient (

[0175] ;

[0176] In the formula, is the predicted value of particulate matter deposition in the drainage pipe network on the th day. Use the observed data of particulate matter deposition in the drainage pipe network and the day's input data for particulate matter deposition prediction in the drainage pipe network (volume sand content under non-deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( )) to predict the in-pipe particulate matter deposition in the drainage pipe network on the th day (i.e., the in-pipe particulate matter deposition in the drainage pipe network in the first step). Then, define the second-step advanced in-pipe particulate matter deposition prediction model in the drainage pipe network as:

[0177] ;

[0178] Using the -day in-pipe particulate matter deposition observation data in the drainage pipe network, -day in-pipe particulate matter deposition prediction data in the drainage pipe network, and -day in-pipe particulate matter deposition prediction input data in the drainage pipe network, predict the in-pipe particulate matter deposition in the drainage pipe network on the th day (i.e., the in-pipe particulate matter deposition in the drainage pipe network in the second step). Similarly, define the th-step advanced in-pipe particulate matter deposition prediction model in the drainage pipe network as:

[0179] ;

[0180] Using the -day in-pipe particulate matter deposition observation data in the drainage pipe network, -day in-pipe particulate matter deposition prediction data in the drainage pipe network, and -day in-pipe particulate matter deposition prediction input data in the drainage pipe network, predict the in-pipe particulate matter deposition situation in the drainage pipe network on the th day (i.e., the in-pipe particulate matter deposition in the drainage pipe network in the th step);

[0181] Step 2.2, use the first model in the step-by-step prediction framework of in-pipe particulate matter deposition in the drainage pipe network to predict the first-step advanced in-pipe particulate matter deposition in the drainage pipe network, and use it as the input for the (second, third,..., m )-step advanced in-pipe particulate matter deposition prediction models in the drainage pipe network (i.e., the second, third,..., m th models in the step-by-step prediction framework of in-pipe particulate matter deposition in the drainage pipe network);

[0182] Step 2.3, use the i (1 < i < m )-th model in the step-by-step prediction framework of in-pipe particulate matter deposition in the drainage pipe network to predict the i -th step advanced in-pipe particulate matter deposition in the drainage pipe network, and use it as the input for the ( i + 1,..., m) Input of the prediction model for particulate matter deposition in the drainage pipe network in advance (i.e., the i +(1)st, …, the m th model in the step prediction framework for particulate matter deposition in the drainage pipe network);

[0183] Step 2.4, use the m th model in the step prediction framework for particulate matter deposition in the drainage pipe network to predict the particulate matter deposition in the drainage pipe network m steps in advance;

[0184] Step 2.5, combining Steps 2.2 - 2.4, through m models to predict the m step prediction data of particulate matter deposition in the drainage pipe network in advance.

[0185] Step 2.5.1, train the first prediction model for particulate matter deposition in the drainage pipe network in advance, and use the volume sand content under non - deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median particle size of sediment ( ), hydraulic radius ( ), and pipe friction coefficient ( ) data to predict the output data of the first - step prediction of particulate matter deposition in the drainage pipe network in advance;

[0186] Step 2.5.2, train the second prediction model for particulate matter deposition in the drainage pipe network in advance, and use the volume sand content under non - deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median particle size of sediment ( ), hydraulic radius ( ), and pipe friction coefficient ( ) data and the output data of the first - step prediction of particulate matter deposition in the drainage pipe network in advance to predict the output data of the second - step prediction of particulate matter deposition in the drainage pipe network in advance;

[0187] Step 2.5.3, train the i ( )th prediction model for particulate matter deposition in the drainage pipe network in advance, and use the volume sand content under non - deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median particle size of sediment ( ), hydraulic radius ( ), and pipe friction coefficient ( ) data and the output data of the particulate matter deposition in the drainage pipe network in advance for the previous i - 1 steps to predict the iOutput data of particulate matter deposition in the advanced drainage pipe network

[0188] Step 2.5.4, summarize the m Output data of particulate matter deposition in the drainage pipe network obtained in steps 2.5.1 - 2.5.3 to obtain m Prediction data of particulate matter deposition in the advanced drainage pipe network

[0189] Step 3, use 5 sets of non - deposition layer data to train the particulate matter deposition prediction model for the drainage pipe network. Each long - short - term memory network unit contains 5 dimensions, including: volume sand content under non - deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( );

[0190] Set the hidden state length of each long - short - term memory network unit to 32. To avoid overfitting, set the dropout rate to 0.1, the learning rate to 0.001, the batchsize (number of samples per batch) to 256, and use the root mean square error (MSE) as the loss function;

[0191] Normalize the input data of the 5 dimensions of each long - short - term memory network unit. All input data are normalized using the z - score normalization method;

[0192] Successively use 5 sets of non - deposition data to train the particulate matter sediment prediction model for the drainage pipe network, and respectively obtain k NSE values under 5 sets of non - deposition data. Select the set of hyperparameters that produces the highest NSE median among all hyperparameter combinations as the hyperparameters of the particulate matter deposition sequence step prediction framework for the drainage pipe network;

[0193] As Figure 8 shown, after training the particulate matter deposition prediction model for the drainage pipe network using the first set of non - deposition data, a comparison diagram of the model prediction data and the real data. As shown in the figure, the Froude number predicted by the model is relatively close to the Froude number of the real observed data. At the same time, the between the prediction data and the real observed data is 0.70;

[0194] As Figure 9 shown, after training the particulate matter deposition prediction model for the drainage pipe network using the first set of non - deposition data, a comparison diagram of the model prediction data and the real data. As shown in the figure, the Froude number predicted by the model is relatively close to the Froude number of the real observed data. At the same time, the between the prediction data and the real observed data is 0.80;

[0195] As Figure 10 shown, after training the prediction model of particulate matter deposition in the drainage pipe network using the first set of non-deposition data, a schematic diagram of the comparison between the model prediction data and the real data. As shown in the figure, the Froude number predicted by the model is relatively close to the Froude number of the real observation data. At the same time, the is 0.843;

[0196] As Figure 11 shown, after training the prediction model of particulate matter deposition in the drainage pipe network using the first set of non-deposition data, a schematic diagram of the comparison between the model prediction data and the real data. As shown in the figure, the Froude number predicted by the model is relatively close to the Froude number of the real observation data. At the same time, the is 0.895;

[0197] As Figure 12 shown, after training the prediction model of particulate matter deposition in the drainage pipe network using the first set of non-deposition data, a schematic diagram of the comparison between the model prediction data and the real data. As shown in the figure, the Froude number predicted by the model is relatively close to the Froude number of the real observation data. At the same time, the is 0.923;

[0198] A double-layer prediction framework for particulate matter deposition sequence in the drainage pipe network is constructed using a long short-term memory network. As Figure 10 shown, this framework includes two encoder sequences, a state vector, and a decoder sequence. The first decoder sequence has long short-term memory network units, and its inputs include the volume sand content under non-deposition conditions of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( ). The second decoder sequence has long short-term memory network units, and its input data is the observed data of particulate matter deposition in the drainage pipe network. The state vector contains the compressed and processed information from the inputs of these two encoder sequences. The decoder sequence has m long short-term memory network units, and the output is the m step-ahead prediction data of particulate matter deposition in the drainage pipe network, as Figure 11 shown;

[0199] The double-layer prediction framework for particulate matter deposition sequence in the drainage pipe network is trained using k actual data of particulate matter deposition in the drainage pipe network. Through k times of training, kFor the NSE values, select the set of hyperparameters that produces the highest median NSE among all hyperparameter combinations as the hyperparameters for the prediction framework of particulate matter deposition in the double-layer drainage pipe network.

[0200] Step 4: Use the step prediction framework for particulate matter deposition in the drainage pipe network to simulate the particulate matter deposition in the drainage pipe network, and calculate the Froude number using the particulate matter deposition prediction benchmark model and empirical formula in the drainage pipe network respectively to evaluate the prediction performance of the step prediction framework for particulate matter deposition in the drainage pipe network.

[0201] Step 4.1: Construct a two-layer long short-term memory network as the particulate matter deposition prediction benchmark model in the drainage pipe network to evaluate the prediction performance of the step prediction framework for particulate matter deposition in the drainage pipe network.

[0202] Step 4.1.1: Use a long short-term memory network to construct a prediction framework for particulate matter deposition sequence in the double-layer drainage pipe network. As Figure 10 shown, this framework consists of two encoder sequences, a state vector, and a decoder sequence. The first decoder sequence has long short-term memory network units, and its inputs include the volume sand content under the non-deposition condition of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( ). The second decoder sequence has long short-term memory network units, and its input data is the observed data of particulate matter deposition in the drainage pipe network. The state vector contains the compressed and processed information of the inputs from these two encoder sequences. The decoder sequence has m long short-term memory network units, and the output is the particulate matter deposition prediction data m steps ahead in the drainage pipe network, as Figure 11 shown;

[0203] Step 4.1.2: Use k actual particulate matter deposition prediction data in the drainage pipe network to train the prediction framework for particulate matter deposition sequence in the double-layer drainage pipe network. Through k times of training, obtain k NSE values, and select the set of hyperparameters that produces the highest median NSE among all hyperparameter combinations as the hyperparameters for the prediction framework of particulate matter deposition sequence in the double-layer drainage pipe network.

[0204] Step 4.2: Through the input data for particulate matter deposition prediction in the drainage pipe network (volume sand content under the non-deposition condition of the sediment bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( )) to predict the particulate matter deposition in the drainage pipe network, and compare it with the output results of the step prediction framework for the particulate matter deposition sequence in the drainage pipe network.

[0205] Step 4.2.1, select j actual observed data of the particulate matter deposition in the drainage pipe network, and input them into the step prediction framework for the particulate matter deposition sequence in the drainage pipe network and the double-layer prediction framework for the particulate matter deposition sequence in the drainage pipe network respectively to obtain the predicted data of the particulate matter deposition in the drainage pipe network;

[0206] Step 4.2.2, after obtaining the predicted data of the particulate matter deposition in the drainage pipe network, calculate the consistency index, mean absolute error, root mean square error, coefficient of determination, and adjustment coefficient, and compare and evaluate the prediction performance of the step prediction framework for the particulate matter deposition sequence in the drainage pipe network and the double-layer prediction framework for the particulate matter deposition sequence in the drainage pipe network according to these indexes.

[0207] Step 4.3, input data (volume sand content under non-deposition conditions of the deposition bed ( ), dimensionless particle diameter ( ), median sediment particle size ( ), hydraulic radius ( ), and pipe friction coefficient ( )) into the particulate matter deposition prediction model in the drainage pipe network to predict the output data of the particulate matter deposition prediction model in the drainage pipe network (Froude number of the particulate matter ( )) and compare it with the Froude number calculated by the empirical formula to analyze the prediction performance of the model.

[0208] As Figure 13 shown, the prediction performance of the prediction model, empirical model, and two empirical formulas using the first set of real observed data is given. As shown in the figure, when comparing and evaluating the prediction performance in the first set of observed data of the particulate matter deposition in the drainage pipe network, the coefficient of determination and adjustment coefficient calculated by the particulate matter deposition prediction model are both better than the results calculated by the empirical formula;

[0209] As Figure 14 shown, the prediction performance of the prediction model, empirical model, and two empirical formulas using the second set of real observed data is given. As shown in the figure, when comparing and evaluating the prediction performance in the second set of observed data of the particulate matter deposition in the drainage pipe network, compared with the benchmark model, the five indexes of the particulate matter deposition prediction model are all better than the benchmark model;

[0210] As Figure 15 shown, the model prediction performances obtained by the prediction model, empirical model, and two empirical formulas using the third set of real observation data are given. As shown in the figure, when comparing and evaluating the model prediction performance in the third set of observed data on particulate matter deposition in the drainage pipe network, the mean absolute error calculated by the particulate matter deposition prediction model is better than the result calculated by the empirical formula;

[0211] As Figure 16 shown, the model prediction performances obtained by the prediction model, empirical model, and two empirical formulas using the fourth set of real observation data are given. As shown in the figure, when comparing and evaluating the model prediction performance in the fourth set of observed data on particulate matter deposition in the drainage pipe network, both the mean absolute error and root mean square error calculated by the particulate matter deposition prediction model are better than the results calculated by the empirical formula;

[0212] As Figure 17 shown, the model prediction performances obtained by the prediction model, empirical model, and two empirical formulas using the fifth set of real observation data are given. As shown in the figure, when comparing and evaluating the model prediction performance in the fifth set of observed data on particulate matter deposition in the drainage pipe network, both the coefficient of determination and adjusted coefficient calculated by the particulate matter deposition prediction model are better than the results calculated by the empirical formula.

Claims

1. A method for predicting particle deposition in a drainage network based on long short-term memory step sequences, characterized in that: The following steps are involved: Step 1, prepare the input data and output data for prediction of particle deposition in the drainage network and preprocess the data; Step 2: construct a prediction model for particle deposition in the drainage pipe network based on the long short-term memory step sequence to predict the particle deposition in the drainage pipe network; Step 3: train and test the model and set the model parameters to ensure that the model can accurately simulate the particle deposition in the drainage network; Step 4, simulate the particle deposition in the drainage network and evaluate the prediction performance of the model; The specific steps of step 1 are: Step 1.1, select the main influencing factors of the prediction of particle deposition in the drainage network, that is, input data, and decision variables, that is, output data; among them, the influencing factors mainly include: volume sand content of the sediment bed under non-deposition conditions , dimensionless particle diameter , median particle size of sediment , hydraulic radius and pipe friction coefficient ; The output data of the prediction model is the Froude number of particles ; Step 1.2, after collecting all the input data and output data of the particle deposition prediction model in the drainage network, check the quality of the data, and use a sliding time window to replace the missing data and erroneous data in the input data and output data to ensure that the quality of the input data and output data of the prediction model meets the prediction requirements; The specific steps of step 2 are: Step 2.1: Construct a step prediction framework for particle deposition sequence in drainage pipe network based on long short-term memory network. The framework includes m Models for implementation m Prediction of particle deposition in drainage pipe network; Step 2.2: Use the first model in the step prediction framework of the particle deposition sequence in the drainage network to predict the particle deposition in the drainage network in the first step, and use it as the second, third, ..., and third step prediction models. m The input of the prediction model of particle deposition in the drainage network in advance is the second, third, ..., and third order in the step prediction framework of the particle deposition sequence in the drainage network. m Models; Step 2.3: Use the first step of the particle deposition sequence prediction framework in the drainage network. i (1< i < m ) model predicts the i The first step is to advance the deposition of particles in the drainage network and serve as the first i +1,…,th m The input of the prediction model of particle deposition in the drainage network in advance is the first step in the step prediction framework of particle deposition sequence in the drainage network. i +1, …, m Models; Step 2.4: Use the first step of the particle deposition sequence prediction framework in the drainage network. m The model predicts the m Step ahead to prevent the deposition of particulate matter in the drainage network; Step 2.5, combined with steps 2.2-2.4, by m The model predicts m Step ahead to predict the particle deposition in the drainage network; The specific steps of step 3 are: Step 3.1, determine the learnable parameters and hyperparameters of the particle deposition prediction model in the drainage network, so as to predict the particle deposition in the drainage network in the step prediction framework. m The model is trained and tested to ensure that the prediction model can accurately output the prediction data of particle deposition in the drainage network; Step 3.2, determine the performance evaluation index of the particle deposition prediction model in the drainage network, ensure that each model of the step prediction framework of particle deposition in the drainage network can be effectively trained, and obtain the particle deposition prediction model in the drainage network that meets the accuracy requirements; Step 3.3, input the input data of the particle deposition prediction model in the drainage pipe network into the particle deposition prediction model framework in the drainage pipe network, train and test the particle deposition prediction model in the drainage pipe network, and ensure that the output data of the particle deposition prediction model in the drainage pipe network can accurately reflect the actual change of particle deposition in the drainage pipe network; The step 4 is specifically as follows: Step 4.1, construct a two-layer long short-term memory network as a benchmark model for predicting particle deposition in drainage pipe networks, and evaluate the prediction performance of the step prediction framework for particle deposition sequence in drainage pipe networks; Step 4.2: Predict the input data of particle deposition in the drainage network: the volumetric sand content of the sedimentary bed under non-deposition conditions , dimensionless particle diameter , median particle size of sediment , hydraulic radius and pipe friction coefficient The predicted time series data of particle deposition in the drainage network are obtained and compared with the output results of the step prediction framework of particle deposition sequence in the drainage network; Step 4.3, input data through the prediction model of particle deposition in the drainage network: volumetric sand content of the sediment bed under non-deposition conditions , dimensionless particle diameter , median particle size of sediment , hydraulic radius and pipe friction coefficient , the output data of the particle deposition prediction model in the drainage network are predicted: the Froude number of the particles and compared it with the Froude number calculated by the empirical formula to analyze the prediction performance of the model; The step 3.1 is specifically as follows: Step 3.1.1, determine the dimensions of each LSTM network unit. Each network unit contains 5 dimensions, including: volume sand content of the sedimentary bed under non-sedimentary conditions , dimensionless particle diameter , median particle size of sediment , hydraulic radius and pipe friction coefficient ; Step 3.1.2, normalize the five-dimensional input data of each LSTM network unit. All input data are normalized using the z-score normalization method; Step 3.1.3, set the hidden state length of each LSTM network unit to 32, set the push rate to 0.1 to avoid overfitting, set the learning rate to 0.001, set the number of samples per batch to 256, and use the root mean square error MSE as the loss function; Step 3.1.4, select k real observation data of particle deposition in the drainage network to train the proposed step prediction framework of particle deposition sequence in the drainage network, obtain k NSE values ​​through k training, and select the hyperparameter set that produces the highest NSE median value among all hyperparameter combinations as the hyperparameter of the step prediction framework of particle deposition sequence in the drainage network; The step 4.2 is specifically as follows: Step 4.2.1, select The actual observation data of particle deposition in the drainage pipe network of each city were input into the step prediction framework of particle deposition sequence in the drainage pipe network and the double-layer drainage pipe network prediction framework to obtain the prediction data of particle deposition in the drainage pipe network. Step 4.2.2, after obtaining the prediction data of particle deposition in the drainage network, the consistency index, mean absolute error, root mean square error, Coefficient of determination and Adjustment coefficients are used, and the prediction performance of the particle deposition sequence step prediction framework in the drainage network and the particle deposition prediction framework in the double-layer drainage network are compared and evaluated based on these indicators.

2. According to claim 1, a method for predicting particle deposition in a drainage network based on a long short-term memory step sequence is characterized in that: The step 2.1 is specifically as follows: Step 2.1.1, construct a step prediction model for the particle deposition sequence in a single drainage network pipe, and obtain the particle deposition prediction output data in the drainage network pipe by predicting the particle deposition prediction input data in the drainage network pipe; Step 2.1.2, Build The step prediction model of particle deposition sequence in the drainage network is obtained, and the step prediction framework of particle deposition sequence in the drainage network is obtained, including A sequence step prediction model for particulate matter in the drainage network is used to predict the particle deposition in the drainage network using the input data. Prediction data of particle deposition in drainage pipe network.

3. The method for predicting particle deposition in a drainage network based on a long short-term memory step sequence according to claim 1 is characterized in that: The step 2.5 is specifically as follows: Step 2.5.1, training the first advance drainage pipe network pipe particle deposition prediction model, and using the drainage pipe network pipe particle deposition prediction input data to predict the first step advance drainage pipe network pipe particle deposition prediction output data; Step 2.5.2, training the second advance drainage pipe network pipe particle deposition prediction model, using the drainage pipe network pipe particle deposition prediction input data and the first step advance drainage pipe network pipe particle deposition prediction output data to predict the second step advance drainage pipe network pipe particle deposition prediction output data; Step 2.5.3, for i ( ) advanced drainage pipe network particle deposition prediction model is trained, using the drainage pipe network particle deposition prediction input data and the previous i -1 step ahead of the prediction output data of the particle deposition in the drainage network to obtain the first i Step ahead to predict the output data of particle deposition in the drainage network; Step 2.5.4, summarize the m-step drainage network pipe particle deposition prediction output data obtained in steps 2.5.1-2.5.3 to obtain the m-step advanced drainage network pipe particle deposition prediction data.

4. The method for predicting particle deposition in a drainage pipe network based on a long short-term memory step sequence according to claim 1 is characterized in that: The step 4.1 is specifically as follows: Step 4.1.1, use the long short-term memory network to build a double-layer drainage pipe network particle deposition sequence prediction framework, which contains two encoder sequences, a state vector and a decoder sequence; the first decoder sequence has a long short-term memory network unit, and its input includes the volume sand content of the sediment bed under non-deposition conditions , dimensionless particle diameter , median particle size of sediment , hydraulic radius and pipe friction coefficient ; The second decoder sequence has long short-term memory network units, and its input data is the observation data of particle deposition in the drainage network pipes; the state vector contains the compressed and processed information of the inputs from these two encoder sequences; The decoder sequence has m LSTM network units, and the output is the m-step-ahead prediction of particle deposition in the sewer network; In step 4.1.2, the actual data of particle deposition in the drainage pipe networks of k cities are used to train the particle deposition sequence prediction framework in the double-layer drainage pipe network. K NSE values ​​are obtained through k training times, and the hyperparameter set that produces the highest NSE median value among all hyperparameter combinations is selected as the hyperparameter of the particle deposition sequence prediction framework in the double-layer drainage pipe network.

Images