Urban rainfall runoff pollution prediction method based on integrated rolling decomposition method and deep learning algorithm

Abstract

The present invention relates to urban rainfall runoff pollution prediction in urban water systems, and provides an urban rainfall runoff pollution prediction method based on integrated rolling decomposition method and deep learning algorithm. A rolling decomposition method is firstly used to decompose rainfall runoff sequence data into different sub-sequences; then decomposition is sequentially performed on added data, and future data is excluded, to prevent information leakage; a recurrent neural network is used to model and predict the sub-sequences; and finally, predicted results of the sub-sequences are summed to obtain the predicted result of rainfall runoff pollution.

Description

Urban rainfall runoff pollution prediction method based on rolling decomposition method and deep learning algorithm integration Technical Field

[0001] The present invention relates to urban rainfall runoff pollution prediction of urban water systems, and in particular to an urban rainfall runoff pollution prediction method based on the integration of a rolling decomposition method and a deep learning algorithm. Background Art

[0002] Urban rainfall runoff pollution simulation is of great significance in urban hydrology and has long attracted widespread attention. Various urban rainfall runoff pollution prediction models have been widely used in urban non-point source pollution. Traditional methods include SWMM, STORM, HSPF, MOUSE, InfoWorks ICM, and WASP. These methods predict the migration and transformation of pollutants during rainfall runoff based on the hydrological and hydraulic characteristics of urban rainfall, pollutant accumulation and scour, soil erosion, and the movement of pollutants in drainage networks. In recent years, data-driven methods such as artificial neural networks, recursive neural networks, and long short-term memory networks (LSTMs) have been widely used in urban water systems. These methods can leverage big data to simulate highly nonlinear and complex systems and have demonstrated excellent performance in the hydrological field.

[0003] To further improve prediction accuracy, existing methods often combine data decomposition with deep learning algorithms to construct decomposition ensemble models. These ensemble models typically first decompose the highly complex prediction sequence, then model and predict relatively simple subsequences, and finally integrate the results of the subsequences to obtain the prediction result. This reduces the learning difficulty of the model and thus improves prediction accuracy.

[0004] However, while decomposition ensemble models may superficially improve prediction accuracy, the one-time decomposition of data often leads to information leakage, which can adversely affect predictions. Information leakage refers to the fact that the decomposition results of data at a specific time are affected by subsequent data available after that time. Therefore, information leakage can lead to falsely high accuracy in models that are ineffective in real-world applications because future information is no longer available.

[0005] In view of this, the present invention proposes an urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm for urban rainfall runoff pollution simulation, which is used for rainfall runoff pollution prediction. Improving urban rainfall runoff pollution prediction is of great significance for the prevention and control of urban water environment pollution. The method adopts variational mode decomposition as the data decomposition algorithm and uses recursive neural network to construct a rainfall runoff pollution prediction model. First, the rolling decomposition method is used to decompose the rainfall runoff sequence data into different subsequences, and data is added in sequence for decomposition, excluding future data to avoid information leakage. The subsequences are modeled and predicted using a recursive neural network. Finally, the prediction results of the subsequences are summed to obtain the prediction results of rainfall runoff pollution.

[0006] Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide an urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm. The rolling decomposition method is used to decompose the rainfall runoff sequence data into different subsequences, and data is added in sequence for decomposition, and future data is excluded to avoid information leakage. A recursive neural network is used to model and predict the subsequences. Finally, the prediction results of the subsequences are summed to obtain the prediction results of rainfall runoff pollution.

[0008] To solve the above technical problems, the technical solution adopted by the present invention is: a method for predicting urban rainfall runoff pollution based on the integration of rolling decomposition method and deep learning algorithm, comprising the following steps:

[0009] S1. Prepare input data and output data for rainfall runoff pollution prediction;

[0010] S2. Decompose and reorganize the original sequence of rainfall runoff pollution using the rolling decomposition method, add rainfall runoff pollution prediction input data in sequence, and exclude future data through sequence decomposition;

[0011] S3. Construct a rainfall runoff pollution prediction model based on the integration of the rolling decomposition method and the deep learning algorithm, and use the Bayesian optimization algorithm to update and optimize the parameters of the rainfall runoff pollution prediction model. Use each rainfall runoff pollution prediction input data subsequence to predict a single runoff pollution prediction result, and fuse the single runoff pollution prediction results to obtain the rainfall runoff pollution prediction output data.

[0012] S4. Determine the evaluation indicators for rainfall runoff pollution prediction performance, use real rainfall runoff pollution observation data to train and test the rainfall runoff pollution prediction model, and evaluate the prediction performance of the rainfall runoff pollution prediction model based on the rainfall runoff pollution prediction results.

[0013] In a preferred embodiment, step S1 includes the following steps:

[0014] S1.1. Collect real-world observational data on rainfall runoff pollution from different cities across China and divide the data into two parts: influencing factors and decision variables based on the main influencing factors and observation indicators of rainfall runoff pollution.

[0015] The influencing factors include: rainfall type, average rainfall intensity, peak rainfall intensity, rainfall duration, dry period before rain, atmospheric dust fall, PM10, vehicle volume, pavement material and functional area. These data are used as input data to predict rainfall runoff pollution data.

[0016] The decision variables mainly include: average concentration of rainfall runoff pollution EMC and initial flushing index FF30, which are used as output variables of the prediction model;

[0017] S1.2. Conduct quality checks on real observation data of rainfall runoff pollution in different cities, remove outliers in the real dataset, and fill in missing values;

[0018] S1.3. Based on the actual attributes of the observed rainfall runoff data, convert the non-numerical rainfall runoff observation variables into corresponding numerical values. The non-numerical rainfall runoff observation variables include: rainfall type, pavement material, and functional area;

[0019] Rainfall patterns are classified by the number of rainfall peaks, which can be categorized as unimodal, bimodal, and multimodal.

[0020] Pavement materials are divided into roads, squares, green spaces and water surfaces;

[0021] Functional areas are divided into industrial areas, commercial areas and residential areas;

[0022] S1.4. Divide the rainfall runoff pollution data into three parts: training data, validation data, and test data, and normalize the rainfall runoff pollution prediction input data.

[0023] In a preferred embodiment, the step S1.2 includes the following steps:

[0024] S1.2.1. Use the Tukey test to identify outliers in the rainfall runoff pollution observation data and treat outliers as missing values;

[0025] S1.2.2 For rainfall runoff pollution observation data, if there are 10 consecutive moments of missing rainfall observation data, the rainfall runoff pollution observation data for that session shall be discarded;

[0026] S1.2.3. For missing values ​​in rainfall runoff pollution observation data, the average value of the observation data at the eight moments before and after the missing value is selected as the rainfall runoff pollution data observation value at that moment.

[0027] In a preferred embodiment, step S2 includes the following steps:

[0028] S2.1. Decompose the original rainfall runoff pollution sequence into multiple subsequences, including six subsequences: U1, U2, U3, U4, U5, and Res;

[0029] S2.2. Reorganize the five subsequences U2, U3, U4, U5 and Res into a new subsequence Usum.

[0030] In a preferred solution, in step S2.1, a data decomposition algorithm is used to decompose the rainfall runoff sequence into subsequences, and data points are added one by one during the decomposition process to perform rolling decomposition on the rainfall runoff sequence.

[0031] In a preferred embodiment, step S3 includes the following steps:

[0032] S3.1. Use the recursive neural network (RGU) to build a rainfall runoff pollution prediction model, and use rainfall observation data to predict runoff pollution data.

[0033] S3.2. Use the rainfall runoff pollution subsequences U1 and Usum to train and test the rainfall runoff pollution prediction model, and use the Bayesian optimization algorithm to optimize the hyperparameters of each independent model of the subsequence;

[0034] S3.3, using the input data of the rainfall runoff pollution subsequence U1 and Usum to predict the runoff pollution output data;

[0035] S3.4. Combine the predicted data of the rainfall runoff pollution subsequences U1 and Usum and fuse them to obtain the final predicted data of the model.

[0036] In a preferred solution, in step S3.1, a recursive neural network is used to construct a rainfall runoff pollution prediction model, and the single rainfall runoff pollution prediction input data is used to predict the average rainfall runoff pollution concentration EMC and the initial flushing index FF30.

[0037] In a preferred embodiment, step S3.2 includes the following steps:

[0038] S3.2.1. Determine the hyperparameters of the recurrent neural network unit for rainfall runoff pollution prediction. Set the number of neural network layers to 3, the number of epochs to 800, select MAE as the loss function, and select Adam as the optimizer.

[0039] S3.2.2. Use the rainfall runoff pollution subsequence U1 to train and test the rainfall runoff pollution prediction model, and use the Bayesian optimization algorithm to optimize the hyperparameters of the recursive neural network unit to obtain the parameter combination with the best runoff pollution prediction performance;

[0040] S3.2.3. Use the rainfall runoff pollution subsequence Usum to train and test the rainfall runoff pollution prediction model, and use the Bayesian optimization algorithm to optimize the hyperparameters of the recursive neural network unit to obtain the parameter combination with the best runoff pollution prediction performance;

[0041] In a preferred embodiment, step S4 includes the following steps:

[0042] S4.1. Using k sets of real rainfall runoff pollution observation data, predict runoff pollution data;

[0043] S4.2. Use the predicted runoff pollution data to evaluate the performance of the rainfall runoff pollution prediction model.

[0044] In a preferred embodiment, in step S4.2, the performance of the rainfall runoff pollution prediction model is evaluated by calculating the mean absolute error, mean absolute percentage error, and root mean square error of the predicted average rainfall runoff pollution concentration EMC and the initial flushing index FF30.

[0045] The mean absolute error is calculated as:

[0046] The mean absolute percentage error is calculated as:

[0047] The formula for calculating the root mean square error is:

[0048] Where, Indicates the predicted value of EMC or FF30, y i represents the EMC or FF30 observation value, and N represents the total number of EMC or FF30.

[0049] The present invention provides a method for predicting urban rainfall runoff pollution based on the integration of a rolling decomposition method and a deep learning algorithm, which has the following beneficial effects:

[0050] 1. This method uses variational mode decomposition as a data decomposition algorithm and a recursive neural network to construct a rainfall runoff pollution prediction model. First, a rolling decomposition method is used to decompose the rainfall runoff series data into different subsequences. Data is added sequentially for decomposition, and future data is excluded to avoid information leakage. A recursive neural network is then used to model and predict the subsequences. Finally, the prediction results of the subsequences are summed to obtain the rainfall runoff pollution prediction result. This can significantly improve the prediction accuracy and the applicability of the method in different environments.

[0051] 2. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm is of great significance for the prevention and control of urban water environment pollution. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0053] Figure 1 is a flow chart of the urban rainfall runoff pollution prediction method based on the integration of the rolling decomposition method and the deep learning algorithm;

[0054] Figure 2 is a schematic diagram of the one-time decomposition of traditional runoff pollution subsequences;

[0055] Figure 3 is a schematic diagram of the rolling decomposition of the runoff pollution subsequence;

[0056] Figure 4 is a schematic diagram of the recursive neural network structure;

[0057] Figure 5 is a comparison of the EMC predicted values ​​and true values ​​obtained using the rainfall runoff pollution prediction model;

[0058] Figure 6 is a comparison of the FF30 predicted value and the true value obtained using the rainfall runoff pollution prediction model; DETAILED DESCRIPTION

[0059] The specific embodiments of the present invention are further described in detail with reference to FIG1 to FIG6 .

[0060] As shown in Figure 1, a method for predicting urban rainfall runoff pollution based on the integration of rolling decomposition method and deep learning algorithm.

[0061] S1. Prepare the input and output data for rainfall runoff pollution prediction in order to train and test the model. Table 1 shows the rainfall runoff pollution data of 14 times in China.

[0062] Table 1 Domestic rainfall runoff pollution data for 14 times

[0063] S1.1 Collect real observation data on rainfall runoff pollution in different cities in China, and divide the real observation data into two parts: main influencing factors and decision variables based on the main influencing factors and observation indicators of rainfall runoff pollution.

[0064] The main influencing factors of rainfall runoff pollution prediction are shown in Table 2.

[0065] Table 2 Main influencing factors of rainfall runoff pollution prediction

[0066] The influencing factors mainly include: rainfall type, average rainfall intensity, peak rainfall intensity, rainfall duration, dry period before rain, atmospheric dust fall, PM10, traffic volume, pavement material and functional area. These observation data will have an impact on rainfall runoff pollution. These data are used as input data to predict rainfall runoff pollution data.

[0067] The decision variables mainly include: average concentration of rainfall runoff pollution EMC and initial flushing index FF30.

[0068] These two indicators can accurately reflect the runoff pollution caused by rainfall events, and these two data are used as output variables of the prediction model;

[0069] S1.2 Perform quality checks on real observation data of rainfall runoff pollution in different cities, remove outliers in the real dataset, and supplement missing values.

[0070] S1.2.1 Use the Tukey test to identify outliers in the rainfall runoff pollution observation data and treat outliers as missing values;

[0071] The Tukey test defined the interquartile range (IQR) as the difference between the upper quartile Q3 and the lower quartile Q1, and defined data greater than Q3 + 1.5 IQR or less than Q1 - 1.5 IQR as outliers.

[0072] S1.2.2 For rainfall runoff pollution observation data of a session, if rainfall observation data are missing for 10 consecutive moments, the rainfall runoff pollution observation data of that session shall be discarded.

[0073] S1.2.3 For missing values ​​in rainfall runoff pollution observation data, the average value of the observation data at the eight moments before and after the missing value is selected as the rainfall runoff pollution data observation value at that moment.

[0074] S1.3 Based on the actual attributes of the observed data in rainfall runoff, convert the non-numerical rainfall runoff observation variables into corresponding numerical values. The non-numerical rainfall runoff observation variables include: rainfall type, pavement material, and functional area.

[0075] Rainfall types are mainly classified by the number of rainfall peaks, and can be divided into unimodal, bimodal and multimodal types. These three rainfall types can be represented by 1, 2 and 3 respectively.

[0076] Pavement materials can be divided into roads, squares, green spaces and water surfaces, represented by 1, 2, 3 and 4 respectively;

[0077] Functional areas are divided into industrial areas, commercial areas and residential areas, represented by 1, 2 and 3 respectively.

[0078] S1.4 divides the rainfall runoff pollution data into three parts: training data, verification data, and test data, and normalizes the input data.

[0079] S2. As shown in Figure 3, the variational mode decomposition (VMD) under the rolling method is used to decompose the original rainfall runoff pollution sequence into different subsequences. Compared with the traditional sequence decomposition method shown in Figure 2, this method adds data sequentially for decomposition, excludes future data, and avoids information leakage.

[0080] S2.1. Decompose the original rainfall runoff pollution sequence into multiple subsequences, using subsequences U1, U2, U3, U4, U5 and Res as six subsequences.

[0081] S2.1.1 uses a data decomposition algorithm to decompose the rainfall runoff sequence into subsequences, and transforms the signal decomposition problem into a constrained variational problem solution process.

[0082] S2.1.2 Add data points one by one during the decomposition process and perform "rolling decomposition" on the rainfall runoff sequence to ensure that the decomposition process is not affected by subsequent data, and decompose the rainfall runoff pollution sequence into six subsequences: U1, U2, U3, U4, U5 and Res.

[0083] S2.2 reorganizes the five subsequences U2, U3, U4, U5 and Res into a new subsequence Usum.

[0084] S3. A recursive neural network (GRU) is used to model and predict the subsequence. The schematic diagram of the recursive neural network structure is shown in Figure 4.

[0085] S3.1 uses a recursive neural network (RGU) to construct a rainfall runoff pollution prediction model, and uses rainfall observation data to predict runoff pollution data.

[0086] Traditional neural network models have local influences, and the output of the model is mainly affected by adjacent input values. At the same time, the later output is easily affected by the earlier input data, that is, the traditional neural network model is not good at capturing the long-term dependencies of the sequence. In fact, in the process of establishing a rainfall runoff pollution prediction model, not every observation data has the same importance. In the process of model learning, if the model can remember more important information and forget information that contributes little or no contribution. The recursive neural network implements this "selective memory" function through the design of two gates, the update gate and the reset gate, thereby alleviating the gradient disappearance of the neural network, so that the model has the function of long-term memory. Among them, the role of the update gate is to determine the degree to which the current input is retained or discarded, and the role of the reset gate is to determine the degree to which the information of the previous hidden layer is retained or discarded;

[0087] A recurrent neural network is a neural network similar to a long short-term memory network, which replaces the hidden block in the RNN with two logic gates and a candidate hidden state. The hidden block of a recurrent neural network can be represented as:

[0088] Reset the gate:

[0089] Update Gate:

[0090] Candidate hidden layer states:

[0091] Hidden layer state:

[0092] Where W xr 、W xz 、W xh ∈R d×h and W hr 、W hz 、W hh ∈R h×h is the weight matrix; b r 、b z 、b h ∈R 1×h is the deviation. gt Represents the reset gate, Z gt represents the update gate, D t represents the input data, and σ represents the hyperparameter of the neural network.

[0093] Update Gate Z gt It is used to capture the long-term dependencies in the time series. At the same time, the reset gate R gt The algorithm learns short-term dependencies in time series by using candidate hidden states. The candidate hidden state represents the influence of the previous hidden state on the current hidden state. If the element in the reset gate is close to 1, the hidden state of the previous hidden state is retained. If the element in the reset gate is close to 0, the hidden state of the previous hidden state is forgotten.

[0094] S3.1.1 Use recursive neural network (RGU) to construct a rainfall runoff pollution prediction model, and use the single rainfall runoff pollution prediction input data to predict the average rainfall runoff pollution concentration EMC and initial flushing index FF30.

[0095] S3.1.2 Design the forget gate and reset gate of the recursive neural network to ensure that each recursive neural network unit can select more important rainfall runoff pollution prediction input data.

[0096] The update gate is mainly used to determine the extent to which the current input data of the rainfall runoff pollution prediction input data is retained or discarded, while the reset gate is used to determine the extent to which the hidden information of the previous layer of the recurrent neural network unit is retained or discarded.

[0097] S3.2 uses the rainfall runoff pollution prediction input data subsequences U1 and Usum to train and test the rainfall runoff pollution prediction model, and uses the Bayesian optimization algorithm to optimize the hyperparameters of each independent model of the subsequence.

[0098] Bayesian optimization is a powerful strategy for finding the extreme value of an objective function that is expensive to evaluate. It is applicable when there is no closed-form expression for the objective function, but observations of that function (which may be noisy) are available at sampled values. It is particularly useful when these computations are expensive, when one cannot obtain derivatives, or when the problem at hand is non-convex.

[0099] Bayesian optimization techniques are among the most efficient methods in terms of the number of function evaluations required. Much of the efficiency stems from Bayesian optimization's ability to incorporate prior beliefs about the problem to help guide sampling, and to balance exploration and exploitation of the search space. It is called Bayesian because it uses the famous "Bayes' theorem", which states (somewhat oversimplified) that the posterior probability P(M / E) of a model (or theory or hypothesis) M given evidence (or data or observations) E is proportional to the likelihood P(E / M) of E given M multiplied by the prior probability P(M) of M: P(M / E) ∝ P(E / M)P(M)

[0100] Within this simple equation lies the key to optimizing the objective function. In Bayesian optimization, the prior represents our belief about the space of possible objective functions. While the cost function is unknown, it is reasonable to assume that there is prior knowledge about some of its properties, such as smoothness, that makes some possible objective functions more plausible than others.

[0101] Define x i is the i-th sample, f(x i ) is the objective function at x i When the cumulative observations When , the prior distribution and likelihood function Combining. If the prior belief is that the target function is very smooth and noise-free, then data with high variance or oscillations should be considered less likely than data that deviate significantly from the mean. Therefore, these can be combined to get the posterior distribution:

[0102] S3.2.1 Determine the hyperparameters of the recursive neural network unit for rainfall runoff pollution prediction. Set the number of neural network layers to 3, the epoch to 800, select MAE as the loss function, and select Adam as the optimizer.

[0103] S3.2.2 Use the rainfall runoff pollution prediction input data subsequence U1 to train and test the rainfall runoff pollution prediction model, and use the Bayesian optimization algorithm to optimize the hyperparameters of the recursive neural network unit to obtain the parameter combination with the best rainfall runoff pollution prediction performance.

[0104] S3.2.3 The rainfall runoff pollution prediction input data subsequence Usum is used to train and test the rainfall runoff pollution prediction model, and the Bayesian optimization algorithm is used to optimize the hyperparameters of the recursive neural network unit to obtain the parameter combination with the best rainfall runoff pollution prediction performance.

[0105] S3.3 uses the input data of the rainfall runoff pollution subsequences U1 and Usum to predict the runoff pollution output data.

[0106] S3.4 combines the predicted data of the rainfall runoff pollution subsequences U1 and Usum to obtain the final predicted data of the model.

[0107] S4. Determine the evaluation indicators for rainfall runoff pollution prediction performance, use real rainfall runoff pollution observation data to train and test the rainfall runoff pollution prediction model, and evaluate the prediction performance of the rainfall runoff pollution prediction model based on the rainfall runoff pollution prediction results.

[0108] S4.1 uses k sets of real rainfall runoff pollution observation data to predict runoff pollution data.

[0109] S4.2 Use the predicted runoff pollution data to evaluate the performance of the rainfall runoff pollution prediction model.

[0110] During the training and testing process of the rainfall runoff pollution prediction model, the predicted average rainfall runoff pollution concentration EMC and initial flushing index FF30 were used to calculate the average absolute error, average absolute percentage error and root mean square error of the average rainfall runoff pollution concentration EMC and initial flushing index FF30, respectively.

[0111] The mean absolute error is calculated as:

[0112] The mean absolute percentage error is calculated as:

[0113] The formula for calculating the root mean square error is:

[0114] Where, Indicates the predicted value of EMC or FF30, y i represents the EMC or FF30 observation value, and N represents the total number of EMC or FF30.

[0115] The optimal parameter combination for the prediction performance indicators of average rainfall runoff pollution concentration (EMC) and initial flushing index (FF30) was selected as the parameter base for the rainfall runoff pollution prediction model. As shown in Figures 5 and 6, the predicted values ​​obtained using the rainfall runoff pollution prediction model are relatively close to the observed values, indicating that the rainfall runoff pollution prediction model can be used to simulate and predict urban rainfall runoff pollution.

Claims

1. A method for predicting urban rainfall runoff pollution based on the integration of rolling decomposition method and deep learning algorithm, characterized in that: The steps include: S1. Prepare input data and output data for rainfall runoff pollution prediction; S2. Decompose and reorganize the original sequence of rainfall runoff pollution using the rolling decomposition method, add rainfall runoff pollution prediction input data in sequence, and exclude future data through sequence decomposition; S3. Construct a rainfall runoff pollution prediction model based on the integration of the rolling decomposition method and the deep learning algorithm, and use the Bayesian optimization algorithm to update and optimize the parameters of the rainfall runoff pollution prediction model. Use each rainfall runoff pollution prediction input data subsequence to predict a single runoff pollution prediction result, and fuse the single runoff pollution prediction results to obtain the rainfall runoff pollution prediction output data. S4. Determine the evaluation indicators for rainfall runoff pollution prediction performance, use real rainfall runoff pollution observation data to train and test the rainfall runoff pollution prediction model, and evaluate the prediction performance of the rainfall runoff pollution prediction model based on the rainfall runoff pollution prediction results.

2. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm according to claim 1 is characterized in that: The step S1 includes the following steps: S1.

1. Collect real-world observational data on rainfall runoff pollution from different cities across China and divide the data into two parts: influencing factors and decision variables based on the main influencing factors and observation indicators of rainfall runoff pollution. The influencing factors include: rainfall type, average rainfall intensity, peak rainfall intensity, rainfall duration, dry period before rain, atmospheric dust fall, PM10, vehicle volume, pavement material and functional area. These data are used as input data to predict rainfall runoff pollution data. The decision variables mainly include: average concentration of rainfall runoff pollution EMC and initial flushing index FF30, which are used as output variables of the prediction model; S1.

2. Conduct quality checks on real observation data of rainfall runoff pollution in different cities, remove outliers in the real dataset, and fill in missing values; S1.

3. Based on the actual attributes of the observed rainfall runoff data, convert the non-numerical rainfall runoff observation variables into corresponding numerical values. The non-numerical rainfall runoff observation variables include: rainfall type, pavement material, and functional area; Rainfall patterns are classified by the number of rainfall peaks, which can be categorized as unimodal, bimodal, and multimodal. Pavement materials are divided into roads, squares, green spaces and water surfaces; Functional areas are divided into industrial areas, commercial areas and residential areas; S1.

4. Divide the rainfall runoff pollution data into three parts: training data, validation data, and test data, and normalize the rainfall runoff pollution prediction input data.

3. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm according to claim 2 is characterized in that: The step S1.2 includes the following steps: S1.2.

1. Use the Tukey test to identify outliers in the rainfall runoff pollution observation data and treat outliers as missing values; S1.2.2 For rainfall runoff pollution observation data, if there are 10 consecutive moments of missing rainfall observation data, the rainfall runoff pollution observation data for that session shall be discarded; S1.2.

3. For missing values ​​in rainfall runoff pollution observation data, the average value of the observation data at the eight moments before and after the missing value is selected as the rainfall runoff pollution data observation value at that moment.

4. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm according to claim 1 is characterized in that: The step S2 includes the following steps: S2.

1. Decompose the original rainfall runoff pollution sequence into multiple subsequences, including six subsequences: U1, U2, U3, U4, U5, and Res; S2.

2. Reorganize the five subsequences U2, U3, U4, U5 and Res into a new subsequence Usum.

5. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm according to claim 4 is characterized in that: In step S2.1, a data decomposition algorithm is used to decompose the rainfall runoff sequence into subsequences, and data points are added one by one during the decomposition process to perform rolling decomposition on the rainfall runoff sequence.

6. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm according to claim 1 is characterized in that: The step S3 includes the following steps: S3.

1. Use the recursive neural network (RGU) to build a rainfall runoff pollution prediction model, and use rainfall observation data to predict runoff pollution data. S3.

2. Use the rainfall runoff pollution subsequences U1 and Usum to train and test the rainfall runoff pollution prediction model, and use the Bayesian optimization algorithm to optimize the hyperparameters of each independent model of the subsequence; S3.3, using the input data of the rainfall runoff pollution subsequence U1 and Usum to predict the runoff pollution output data; S3.

4. Combine the predicted data of the rainfall runoff pollution subsequences U1 and Usum and fuse them to obtain the final predicted data of the model.

7. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm according to claim 6 is characterized in that: In step S3.1, a rainfall runoff pollution prediction model is constructed using a recursive neural network, and the single rainfall runoff pollution prediction input data is used to predict the average rainfall runoff pollution concentration EMC and the initial flushing index FF30.

8. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm according to claim 6 is characterized in that: The step S3.2 includes the following steps: S3.2.

1. Determine the hyperparameters of the recurrent neural network unit for rainfall runoff pollution prediction. Set the number of neural network layers to 3, the number of epochs to 800, select MAE as the loss function, and select Adam as the optimizer. S3.2.

2. Use the rainfall runoff pollution subsequence U1 to train and test the rainfall runoff pollution prediction model, and use the Bayesian optimization algorithm to optimize the hyperparameters of the recursive neural network unit to obtain the parameter combination with the best runoff pollution prediction performance; S3.2.

3. Use the rainfall runoff pollution subsequence Usum to train and test the rainfall runoff pollution prediction model, and use the Bayesian optimization algorithm to optimize the hyperparameters of the recursive neural network unit to obtain the parameter combination with the best runoff pollution prediction performance.

9. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm according to claim 1 is characterized in that: The step S4 includes the following steps: S4.

1. Using k sets of real rainfall runoff pollution observation data, predict runoff pollution data; S4.

2. Use the predicted runoff pollution data to evaluate the performance of the rainfall runoff pollution prediction model.

10. The urban rainfall runoff pollution prediction method based on the integration of rolling decomposition method and deep learning algorithm according to claim 9 is characterized in that: In step S4.2, the performance of the rainfall runoff pollution prediction model is evaluated by calculating the mean absolute error, mean absolute percentage error, and root mean square error of the predicted rainfall runoff pollution average concentration EMC and initial flushing index FF30. The mean absolute error is calculated as: The mean absolute percentage error is calculated as: The formula for calculating the root mean square error is: Where, Indicates the predicted value of EMC or FF30, y i represents the EMC or FF30 observation value, and N represents the total number of EMC or FF30.

Images