A pump station post-pool water level prediction method and system based on multi-dimensional sequence data correlation analysis

Abstract

The application provides a pump station post-pool water level prediction method and system based on multi-dimensional sequence data correlation analysis. The method comprises: data preprocessing of pump station monitoring data to obtain trend sequence, periodic sequence and residual sequence; prediction of the trend sequence, the encoder is composed of multiple composite convolution layers, the trend sequence is subjected to feature extraction, a parallel neural network with a weighting mechanism is introduced to learn the relationship between different positions in the feature information, the feature information is subjected to weighting processing, the decoder is composed of a recurrent neural network layer with a gating mechanism to decode the weighted information and perform prediction; the periodic sequence is predicted based on a multi-layer perceptron, and the trend sequence prediction value and the periodic sequence prediction value are added to the residual sequence value to obtain the final water level prediction. The method can pay attention to different parts of the input sequence, understand the change rule of the data, deeply mine the complex correlation relationship between the data, and accurately predict the water level of the pump station.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of water level prediction, and particularly relates to a pump station back pool water level prediction method and system based on multi-dimensional sequence data correlation analysis. BACKGROUND

[0002] In modern society, effective management and rational utilization of water resources are of great importance. Accurate prediction of pump station water level plays a key role and has a profound impact on many aspects. Through the prediction of pump station water level, the urban water supply plan can be scientifically and reasonably arranged, the fair and reasonable allocation of water resources can be realized, the demand of various fields can be met, and thus the social and economic development can be promoted. As a key link of water resources management, scientific and reasonable water resources scheduling and management can help to reasonably regulate reservoir storage capacity, maintain the balance of river ecosystem, improve the operation efficiency and safety of water conservancy projects, and provide reliable support for social and economic development and ecological environment protection.

[0003] Water level prediction faces many challenges, including model complexity, large amount of data, multiple dimensions of data, and accurate capture of various influencing factors. Traditional prediction methods often fail to fully consider the complexity and diversity of water level changes, resulting in low accuracy and stability of prediction results. These methods often ignore the complex interaction between multiple factors, affecting the accuracy and stability of the prediction results. Due to the complex correlation between pump station monitoring data, a more comprehensive method is needed to handle it. SUMMARY

[0004] To solve the above technical problems, the application provides a pump station back pool water level prediction method based on multi-dimensional sequence data correlation analysis to solve the above technical problems.

[0005] The first aspect of the application discloses a pump station back pool water level prediction method based on multi-dimensional sequence data correlation analysis, which comprises the following steps:

[0006] Step S1, collecting pump station operation monitoring data and constructing a pump station operation monitoring data set; based on an STL time series decomposition model, performing local weighted polynomial regression fitting on the pump station operation monitoring data set to obtain a decomposition into a trend sequence, a periodic sequence and a residual sequence;

[0007] Step S2, constructing a trend time series sample by sliding window for the trend sequence; using an encoder with three layers of compound convolution layers to extract features from the trend time series sample to obtain encoded features;

[0008] Step S3, a plurality of position weight information of the coding feature is extracted using a plurality of parallel neural networks based on a feature weight mechanism, and the plurality of position weight information is spliced; the spliced weight information is multiplied by a training parameter matrix to obtain a weighted feature;

[0009] Step S4, a decoder based on a recurrent neural network layer with a gating mechanism is used to decode the weighted feature to obtain a prediction sequence of the trend sequence;

[0010] Step S5, a multi-layer fully connected neural network is applied to predict the periodic sequence to obtain a prediction sequence of the periodic sequence; the prediction sequence of the trend sequence, the prediction sequence of the periodic sequence and the residual sequence are summed to obtain a prediction sequence of the pump station back pool water level.

[0011] According to the method of the first aspect of the application, in the step S1, the pump station operation monitoring data includes:

[0012] The boot time, voltage, current, water pump blade angle, pipeline pressure, water pump frequency, cumulative flow, reverse cumulative flow, active power, reactive power, vibration and swing of the unit, outlet pressure, water pump speed, inlet pool water level, front pool water level and back pool water level.

[0013] According to the method of the first aspect of the application, in the step S2, the method for extracting features from the trend time sequence samples using the three-layer composite convolution layer includes:

[0014] The trend time sequence samples are input into the convolution layer with weight value normalization of the first composite convolution layer in series to obtain a weight value normalized convolution output;

[0015] A residual structure is introduced, a lower 1×1 convolution layer is used to adjust the dimension of the trend time sequence samples, so that the dimension is the same as that of the weight value normalized convolution output; after the trend time sequence samples with adjusted dimension are connected with the weight value normalized convolution output, an activation function is input to obtain the first composite convolution output;

[0016] The first composite convolution output is input into the convolution layer with a hole of the second composite convolution layer in series to obtain a hole convolution output;

[0017] After the first composite convolution output is connected with the hole convolution output, an activation function is input to obtain the second composite convolution output;

[0018] The second composite convolution output is input into the third composite convolution layer to obtain a third composite convolution output; the third composite convolution layer is the convolution layer with a hole of the second composite convolution layer in series;

[0019] After connecting the second composite convolution output and the third composite convolution output, an activation function is input to obtain an encoding feature.

[0020] According to the method of the first aspect of the application, in the step S2, the method of inputting the trend time series sample into the convolution layer with weight value normalization of the first composite convolution layer in series to obtain the weight value normalized convolution output comprises:

[0021] The structure of the convolution layer with weight value normalization of the first composite convolution layer is:

[0022]

[0023] After ReLU function activation and Dropout random inactivation, nonlinearity is introduced and overfitting risk is reduced to obtain the output of the convolution layer with weight value normalization of the activated and inactivated first composite convolution layer

[0024] wherein, represents the process output of the convolution layer with weight value normalization of the first composite convolution layer; is the weight of the convolution kernel; b (1) is the bias term of the convolution kernel; c is the size of the convolution kernel; the trend time series sample T τ =[T τ,1 ,T τ,2 ,…,T τ,n ];

[0025]

[0026] wherein, F1() represents the convolution layer with weight value normalization of the first composite convolution layer; represents the weight value normalized convolution output.

[0027] According to the method of the first aspect of the application, in the step S2, the method of inputting the first composite convolution output into the convolution layer with holes of the second composite convolution layer in series to obtain the hole convolution output comprises:

[0028] The structure of the convolution layer with holes of the second composite convolution layer is:

[0029]

[0030] After ReLU function activation and Dropout random inactivation, nonlinearity is introduced and overfitting risk is reduced to obtain the output of the convolution layer with holes of the activated and inactivated second composite convolution layer wherein, represents the output of the convolutional layer with holes of the second composite convolutional layer; f() represents a convolution kernel, d represents an inflation factor, and τ-d·i represents a past direction, represents the convolution of the input data and a convolution kernel with an inflation factor d;

[0031]

[0032] wherein F2() represents the convolutional layer with holes of the second composite convolutional layer; represents the output of the convolutional layer with holes.

[0033] According to the method of the first aspect of the application, in the step S3, the method for extracting the position weighting information of the encoded features using the neural network based on the feature weighting mechanism comprises:

[0034] Q = w Q Y t

[0035] K = w K Y t

[0036] V = w V Y t

[0037] wherein Q represents a query vector, K represents a key vector, and V represents a value vector; w Q , w K , and w V are weight matrices, and the weight matrices of each neural network based on the feature weighting mechanism are different; Y t represents an encoded feature;

[0038] The attention degree W of each position is obtained through the query vector Q and the key vector K,

[0039]

[0040] wherein K T is the transpose of K, d k represents the dimension of the key vector K; and softmax() is a normalized exponential function;

[0041] A i (Q,K,V) = W V

[0042] wherein A i (Q,K,V) represents the position weighting information extracted by the i-th neural network based on the feature weighting mechanism.

[0043] According to the method of the first aspect of the application, in the step S4, the method for decoding the weighted features using the decoder based on the recurrent neural network layer with the gating mechanism comprises:

[0044] The weighted feature selects the forgetting information through a forgetting gate:

[0045] f t = σ(W f [h t-1 , x t ] + b f

[0046] where f t represents the forgetting information; h t-1 represents the output at the previous time; x t represents the weighted feature; σ(·) represents an activation function, W f is a forgetting gate weight matrix, and b f is a forgetting gate bias term;

[0047] The weighted feature determines the update quantity through an input gate:

[0048] i t = σ(W i [h t-1 , x t ] + b i )

[0049] where i t represents the update quantity; W i represents an input gate weight matrix; and b i represents an input gate bias term;

[0050] The weighted feature generates a candidate value of the cell state at the current time through a tanh layer;

[0051]

[0052] where c represents the candidate value of the cell state at the current time; tanh() represents a tanh function; W c is a weight matrix connecting h t-1 at the previous time and x t at the current time; and b c represents a bias term of the tanh layer;

[0053] The cell state at the current time is calculated according to the forgetting information, the update quantity, and the candidate value vector:

[0054]

[0055] where C t represents the cell state at the current time; and C t-1 represents the cell state at the previous time;

[0056] The weighted feature is calculated by an output gate to obtain output gate information:

[0057] o t =σ(W0[h t-1 ,x t ]+b0)

[0058] wherein o t represents the output gate information; W0 represents an output gate weight matrix; and b0 represents an output gate bias term.

[0059] According to the output gate information and the cell state at the current moment, a weighted feature decoded feature is calculated:

[0060] h t =o t tanh(C t )

[0061] wherein h t represents the weighted feature decoded feature.

[0062] The second aspect of the present application discloses a pump station post-pool water level prediction system based on multi-dimensional sequence data correlation analysis, the system comprises:

[0063] The first processing module is configured to collect pump station operation monitoring data and construct a pump station operation monitoring data set; based on an STL time sequence decomposition model, locally weighted polynomial regression fitting is performed on the pump station operation monitoring data set to obtain a decomposition into a trend sequence, a periodic sequence and a residual sequence;

[0064] The second processing module is configured to construct a trend time sequence sample by a sliding window method for the trend sequence; an encoder with three layers of compound convolution layers is used to extract features from the trend time sequence sample to obtain encoded features;

[0065] The third processing module is configured to use multiple parallel neural networks based on a feature weighting mechanism to extract multiple position weighting information of the encoded features, and to splice the multiple position weighting information; the spliced weighted information is multiplied by a training parameter matrix to obtain weighted features;

[0066] The fourth processing module is configured to decode the weighted features based on a decoder with a gating mechanism recurrent neural network layer to obtain a predicted sequence of the trend sequence;

[0067] The fifth processing module is configured to apply a multi-layer fully connected neural network to predict the periodic sequence to obtain a predicted sequence of the periodic sequence; the predicted sequence of the trend sequence, the predicted sequence of the periodic sequence and the residual sequence are summed to obtain a pump station post-pool water level prediction sequence.

[0068] The third aspect of the present application discloses an electronic device. The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for predicting water level of a post-pump station pool based on correlation analysis of multi-dimensional sequence data in any one of the first aspect of the present application.

[0069] The fourth aspect of the present application discloses a computer readable storage medium. The computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the method for predicting water level of a post-pump station pool based on correlation analysis of multi-dimensional sequence data in any one of the first aspect of the present application.

[0070] In summary, the scheme proposed in the present application can focus on different parts of the input sequence, understand the change rule of the data, and thus improve the understanding ability of the model to the input and the prediction accuracy of the model. The model can comprehensively understand the complex structure of time series data, has strong ability in fitting complex relationship, decision-making and prediction under uncertainty, has the ability to learn nonlinear structure and approximate complex functions, is robust to noise and abnormal data, and has better generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0071] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the drawings needed in the specific embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0072] Figure 1 The flow chart of the method for predicting water level of a post-pump station pool based on correlation analysis of multi-dimensional sequence data according to an embodiment of the present application;

[0073] Figure 2 The structural diagram of the system for predicting water level of a post-pump station pool based on correlation analysis of multi-dimensional sequence data according to an embodiment of the present application;

[0074] Figure 3 The structural diagram of the electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0076] The first aspect of the present application discloses a pump station post-pool water level prediction method based on multi-dimensional sequence data correlation analysis. Figure 1 The flow chart of the pump station post-pool water level prediction method based on multi-dimensional sequence data correlation analysis according to an embodiment of the present application is shown in Figure 1 The method comprises the following steps:

[0077] Step S1, collecting pump station operation monitoring data to construct a pump station operation monitoring data set; based on an STL time series decomposition model, performing local weighted polynomial regression fitting on the pump station operation monitoring data set to obtain a decomposition into a trend sequence, a periodic sequence and a residual sequence;

[0078] Step S2, constructing a trend time series sample by a sliding window method for the trend sequence; using an encoder of a three-layer compound convolution layer to extract features of the trend time series sample to obtain encoded features;

[0079] Step S3, using a plurality of parallel neural networks based on a feature weighting mechanism to extract a plurality of position weighting information of the encoded features, and splicing the plurality of position weighting information; multiplying the spliced weighting information with a training parameter matrix to obtain weighted features;

[0080] Step S4, decoding the weighted features based on a decoder of a recurrent neural network layer with a gating mechanism to obtain a prediction sequence of the trend sequence;

[0081] Step S5, applying a multi-layer fully connected neural network to predict the periodic sequence to obtain a prediction sequence of the periodic sequence; summing the prediction sequence of the trend sequence, the prediction sequence of the periodic sequence and the residual sequence to obtain a pump station post-pool water level prediction sequence.

[0082] In step S1, pump station operation monitoring data is collected to construct a pump station operation monitoring data set; based on an STL time series decomposition model, local weighted polynomial regression fitting is performed on the pump station operation monitoring data set to obtain a decomposition into a trend sequence, a periodic sequence and a residual sequence.

[0083] Z t =T t +S t +R t

[0084] wherein Z t is the pump station operation monitoring data set, T t is the trend sequence, S t is the periodic sequence, and R t is the residual sequence.

[0085] In some embodiments, in the step S1, the pump station operation monitoring data comprises:

[0086] The start-up time, voltage, current, water pump blade angle, pipeline pressure, water pump frequency, cumulative flow, reverse cumulative flow, active power, reactive power, vibration and swing of the unit, outlet pressure, water pump speed, inlet pool water level, front pool water level and rear pool water level.

[0087] Specifically, invalid data in the monitoring data is deleted, invalid data is filled using a forward filling method, and a pump station operation monitoring data set is constructed.

[0088] In step S2, a trend time series sample is constructed by sliding window for the trend sequence; an encoder of three-layer composite convolution layers is used to extract features of the trend time series sample to obtain encoded features.

[0089] In some embodiments, in the step S2, the method of using an encoder of three-layer composite convolution layers to extract features of the trend time series sample to obtain encoded features comprises:

[0090] The trend time series sample is input into a convolution layer with weight normalization of a first composite convolution layer in series to obtain a weight normalization convolution output;

[0091] A residual structure is introduced, a lower 1×1 convolution layer is used to adjust the dimension of the trend time series sample so that the dimension is the same as that of the weight normalization convolution output; after the trend time series sample with adjusted dimension is connected with the weight normalization convolution output, an activation function is input to obtain a first composite convolution output;

[0092]

[0093] wherein, the first composite convolution output; Activation() represents an activation function;

[0094] The first composite convolution output is input into a convolution layer with a hole of a second composite convolution layer in series to obtain a hole convolution output;

[0095] After the first composite convolution output is connected with the hole convolution output, an activation function is input to obtain a second composite convolution output;

[0096]

[0097] wherein, the second composite convolution output; Activation() represents an activation function;

[0098] inputting the second composite convolution output into a third composite convolution layer with a convolution layer with a hole in series to obtain a third composite convolution output;

[0099]

[0100] wherein, represents the third composite convolution output;

[0101] connecting the second composite convolution output and the third composite convolution output, and inputting the same into an activation function to obtain an encoding feature;

[0102]

[0103] wherein, Y t represents the encoding feature.

[0104] The method for inputting the trend time sequence sample into the convolution layer with weight value normalization of the first composite convolution layer in series to obtain the weight value normalization convolution output comprises:

[0105] The structure of the convolution layer with weight value normalization of the first composite convolution layer is:

[0106]

[0107] After ReLU function activation and Dropout random inactivation, nonlinearity is introduced and overfitting risk is reduced to obtain the output of the convolution layer with weight value normalization of the first composite convolution layer after activation and inactivation

[0108] wherein, represents the process output of the convolution layer with weight value normalization of the first composite convolution layer; is the weight of the convolution kernel; b (1) is the bias term of the convolution kernel; c is the size of the convolution kernel; the trend time sequence sample T τ =[T τ,1 ,T τ,2 ,...,T τ,n ];

[0109]

[0110] wherein, F1() represents the convolution layer with weight value normalization of the first composite convolution layer; represents the weight value normalization convolution output.

[0111] The method for inputting the first composite convolution output into the convolution layer with a hole of the second composite convolution layer in series to obtain the hole convolution output comprises:

[0112] The structure of the convolution layer with holes of the second composite convolution layer is:

[0113]

[0114] After ReLU function activation and Dropout random inactivation, nonlinearity is introduced and the risk of overfitting is reduced, and the output of the activated and inactivated convolution layer with holes of the second composite convolution layer is obtained Wherein, represents the output of the convolution layer with holes of the second composite convolution layer; f() represents the convolution kernel, d represents the expansion factor, τ-d·i represents the past direction, represents the convolution of the input data and the convolution kernel with the expansion factor d;

[0115]

[0116] Wherein, F2() represents the convolution layer with holes of the second composite convolution layer; represents the output of the convolution layer with holes.

[0117] In step S3, a plurality of parallel neural networks based on feature weighting mechanism are used to extract a plurality of position weighted information of the encoded features, and the plurality of position weighted information are spliced; the spliced weighted information is multiplied by a training parameter matrix to obtain weighted features;

[0118] A = Concat (A1, A2,..., A i ,…,A m )

[0119] x t = AW O

[0120] Wherein, A represents the spliced weighted information; Concat() represents splicing; m represents the number of neural networks based on feature weighting mechanism, A i is a simplified form of A i (Q, K, V); W O represents the training parameter matrix; x t represents the weighted features.

[0121] The method for extracting position weighted information of the encoded features using neural networks based on feature weighting mechanism comprises:

[0122] Q = w Q Y t

[0123] K = w K Y t

[0124] V = w V Y t

[0125] wherein Q represents a query vector, K represents a key vector, and V represents a value vector; w Q , w K , and w V are weight matrices, each of which is different for each feature weighting mechanism-based neural network;

[0126] The attention degree W of each position is obtained through the query vector Q and the key vector K,

[0127]

[0128] wherein K T is the transpose of K, d k represents the dimension of the key vector K; and softmax() is a normalized exponential function;

[0129] A i (Q, K, V) = W V

[0130] wherein A i (Q, K, V) represents the position weighting information extracted by the i-th feature weighting mechanism-based neural network.

[0131] In step S4, a decoder based on a recurrent neural network layer with a gating mechanism is used to decode the weighted features to obtain a predicted sequence of the trend sequence.

[0132] In some embodiments, in the step S4, the method of decoding the weighted features by the decoder based on the recurrent neural network layer with the gating mechanism comprises:

[0133] The weighted features select the forgetting information through a forgetting gate:

[0134] f t = σ(W f [h t-1 , x t ] + b f

[0135] wherein f t represents the forgetting information; h t-1 represents the output at the previous time; x t represents the weighted features; σ(·) represents an activation function; W f is a forgetting gate weight matrix; and b f is a forgetting gate bias term;

[0136] The weighted features determine the update amount through an input gate:

[0137] it = σ(W i [ h t-1 , x t ] + b i )

[0138] wherein i t represents an update amount; W i represents an input gate weight matrix; b i represents an input gate bias term;

[0139] The weighted feature generates a candidate value of a cell state at a current moment through a tanh layer;

[0140]

[0141] wherein h represents the candidate value of the cell state at the current moment; tanh() represents a tanh function; W c is a weight matrix connecting a cell state h t-1 at a previous moment and a current moment x t ; b c represents a bias term of the tanh layer;

[0142] The cell state at the current moment is calculated according to the forgetting information, the update amount, and the candidate value vector:

[0143]

[0144] wherein C t represents the cell state at the current moment; C t-1 represents the cell state at the previous moment;

[0145] The weighted feature calculates output gate information through an output gate:

[0146] o t = σ(W0[h t-1 , x t ] + b0)

[0147] wherein o t represents the output gate information; W0represents an output gate weight matrix; b0represents an output gate bias term;

[0148] The weighted feature decodes the feature according to the output gate information and the cell state at the current moment:

[0149] h t = o t tanh(C t )

[0150] wherein h t represents the weighted feature decoded feature; h tAfter a linear layer, the prediction sequence of the trend sequence is obtained; the cell state of the current time is output to the structure between-1 and 1 through the tanh() function.

[0151] In step S5, the multi-layer fully connected neural network is applied to predict the periodic sequence to obtain the prediction sequence of the periodic sequence; the prediction sequence of the trend sequence, the prediction sequence of the periodic sequence and the residual sequence are summed to obtain the prediction sequence of the water level of the pool behind the pump station.

[0152] Specifically, the multi-layer fully connected neural network is applied to predict the periodic sequence to obtain the prediction sequence of the periodic sequence:

[0153] S predicted =F3(S t )

[0154] Wherein, S predicted represents the prediction sequence of the periodic sequence; F3() represents the multi-layer fully connected neural network.

[0155] The prediction sequence of the trend sequence, the prediction sequence of the periodic sequence and the residual sequence are summed to obtain the prediction sequence of the water level of the pool behind the pump station:

[0156] y predicted =T predicted +S predicted +R t

[0157] Wherein, y predicted represents the prediction sequence of the water level of the pool behind the pump station; T predicted represents the prediction sequence of the trend sequence;

[0158] According to the error between the predicted water level value and the actual water level, the model is updated in gradient to improve the prediction accuracy of the neural network:

[0159] E=∑(y predicted -y true ) 2

[0160]

[0161] Wherein, y true represents the actual water level; E represents the sum of squares of errors; W new is the updated parameter, W old is the current parameter, η is the learning rate, is the partial derivative of the error function with respect to the current parameter. The model can automatically adjust the parameters in the network according to the error between the prediction result and the actual result of each training sample.

[0162] In conclusion, the scheme provided in the application can pay attention to different parts in the input sequence, understand the change rule of data, and thus improve the understanding ability of the model to the input and the prediction accuracy of the model. The model can comprehensively understand the complex structure of the time series data, has strong ability in fitting complex relationship, decision and prediction under uncertainty, has the ability of learning nonlinear structure and approximating complex function, is robust to noise and abnormal data, and has better generalization ability.

[0163] The second aspect of the application discloses a pump station back pool water level prediction system based on multi-dimensional sequence data correlation analysis. Figure 2 For a pump station back pool water level prediction system based on multi-dimensional sequence data correlation analysis according to an embodiment of the application, a structural diagram is shown in FIG. Figure 2 As shown in the figure, the system 100 comprises:

[0164] The first processing module 101 is configured to collect pump station operation monitoring data and construct a pump station operation monitoring data set; based on an STL time series decomposition model, locally weighted polynomial regression fitting is performed on the pump station operation monitoring data set to obtain a decomposition into a trend sequence, a periodic sequence and a residual sequence.

[0165] The second processing module 102 is configured to construct a trend time series sample by a sliding window method for the trend sequence; an encoder with three layers of compound convolution layers is used to extract features from the trend time series sample to obtain encoded features.

[0166] The third processing module 103 is configured to use a plurality of parallel neural networks based on a feature weighting mechanism to extract a plurality of position weighting information of the encoded features, and to splice the plurality of position weighting information; the spliced weighting information is multiplied by a training parameter matrix to obtain weighted features.

[0167] The fourth processing module 104 is configured to decode the weighted features based on a decoder with a recurrent neural network layer with a gating mechanism to obtain a prediction sequence of the trend sequence.

[0168] The fifth processing module 105 is configured to apply a multi-layer fully connected neural network to predict the periodic sequence to obtain a prediction sequence of the periodic sequence; the prediction sequence of the trend sequence, the prediction sequence of the periodic sequence and the residual sequence are summed to obtain a pump station back pool water level prediction sequence.

[0169] According to the system of the second aspect of the application, the first processing module 101 is specifically configured to collect the pump station operation monitoring data, which comprises:

[0170] The start-up time, voltage, current, water pump blade angle, pipeline pressure, water pump frequency, cumulative flow, reverse cumulative flow, active power, reactive power, vibration and swing of the unit, outlet pressure, water pump speed, inlet pool water level, front pool water level and rear pool water level.

[0171] Specifically, invalid data in the monitoring data is deleted, invalid data is filled using a forward filling method, and a pump station operation monitoring data set is constructed.

[0172] According to the system of the second aspect of the application, the second processing module 102 is specifically configured to use an encoder with three composite convolution layers to extract features of the trend time series sample, and the method for obtaining the encoded features comprises:

[0173] The trend time series sample is input into a convolution layer with weight value normalization of a first composite convolution layer in series, to obtain a weight value normalized convolution output;

[0174] A residual structure is introduced, a lower 1×1 convolution layer is used to adjust the dimension of the trend time series sample, so that the dimension is the same as that of the weight value normalized convolution output; after the trend time series sample with the adjusted dimension is connected with the weight value normalized convolution output, an activation function is input, to obtain a first composite convolution output;

[0175]

[0176] wherein, the first composite convolution output is represented by Y1; and Activation() represents an activation function;

[0177] The first composite convolution output is input into a convolution layer with a hole of a second composite convolution layer in series, to obtain a hole convolution output;

[0178] After the first composite convolution output is connected with the hole convolution output, an activation function is input, to obtain a second composite convolution output;

[0179]

[0180] wherein, the second composite convolution output is represented by Y2; and Activation() represents an activation function;

[0181] The second composite convolution output is input into a third composite convolution layer, to obtain a third composite convolution output; the third composite convolution layer is a convolution layer with a hole of a second composite convolution layer in series;

[0182]

[0183] wherein, the third composite convolution output is represented by Y3.

[0184] After connecting the second composite convolution output and the third composite convolution output, an activation function is input to obtain an encoding feature;

[0185]

[0186] Y t represents an encoding feature.

[0187] The method of inputting the trend time sequence sample into the convolution layer with weight value normalization of the first composite convolution layer in series to obtain a weight value normalized convolution output comprises:

[0188] The structure of the convolution layer with weight value normalization of the first composite convolution layer is:

[0189]

[0190] After ReLU function activation and Dropout random inactivation, nonlinearity is introduced and overfitting risk is reduced to obtain the output of the convolution layer with weight value normalization of the first composite convolution layer after activation and inactivation

[0191] wherein, represents the process output of the convolution layer with weight value normalization of the first composite convolution layer; is the weight of the convolution kernel; b (1) is the bias term of the convolution kernel; c is the size of the convolution kernel; the trend time sequence sample T τ =[T τ,1 ,T τ,2 ,··.,T τ,n ];

[0192]

[0193] wherein, F1() represents the convolution layer with weight value normalization of the first composite convolution layer; represents the weight value normalized convolution output.

[0194] The method of inputting the first composite convolution output into the convolution layer with holes of the second composite convolution layer in series to obtain a hole convolution output comprises:

[0195] The structure of the convolution layer with holes of the second composite convolution layer is:

[0196]

[0197] After ReLU function activation and Dropout random inactivation, nonlinearity and overfitting risk reduction are introduced, and the output of the second compound convolution layer with a convolution layer with a hole after activation and inactivation wherein, represents the output of the convolution layer with a hole of the second compound convolution layer; f() represents a convolution kernel, d represents an expansion factor, and tau-d*i represents a past direction, represents the convolution of the input data and the convolution kernel with an expansion factor d;

[0198]

[0199] wherein, F2() represents the convolution layer with a hole of the second compound convolution layer; represents the output of the convolution layer with a hole.

[0200] According to the system of the second aspect of the application, the third processing module 103 is specifically configured to extract position weighting information of the encoded features using a feature weighting mechanism-based neural network:

[0201] Q = w Q Y t

[0202] K = w K Y t

[0203] V = w V Y t

[0204] wherein, Q represents a query vector, K represents a key vector, and V represents a value vector; w Q , w K , and w V are weight matrices, and the weight matrices of each feature weighting mechanism-based neural network are different;

[0205] The attention degree W of each position is obtained through the query vector Q and the key vector K,

[0206]

[0207] wherein, K T is the transpose of K, d k represents the dimension of the key vector K; and softmax() is a normalized exponential function;

[0208] A i (Q,K,V) = WV

[0209] wherein, A i (Q,K,V) represents the position weighting information extracted by the i-th feature weighting mechanism-based neural network.

[0210] According to the system of the second aspect of the present application, the fourth processing module 104 is specifically configured to decode the weighted feature by a decoder based on a recurrent neural network layer with a gating mechanism:

[0211] The weighted feature selects the forgetting information through a forgetting gate:

[0212] f t = σ(W f [h t-1 ,x t ]+b f

[0213] wherein f t represents the forgetting information; h t-1 represents the output of the previous time; x t represents the weighted feature; σ(.) represents an activation function; W f is a forgetting gate weight matrix; and b f is a forgetting gate bias term;

[0214] The weighted feature determines the update amount through an input gate:

[0215] i t = σ(W i [h t-1 ,x t ]+b i )

[0216] wherein i t represents the update amount; W i represents an input gate weight matrix; and b i represents an input gate bias term;

[0217] The weighted feature generates a candidate value of the cell state of the current time through a tanh layer:

[0218]

[0219] wherein represents the candidate value of the cell state of the current time; tanh() represents a tanh function; W c is a weight matrix connecting the previous time h t-1 and the current time x t ; and b c represents a bias term of the tanh layer;

[0220] The cell state of the current time is calculated according to the forgetting information, the update amount, and the candidate value vector:

[0221]

[0222] wherein C ta cell state representing a current time point; C t-1 a cell state representing a previous time point;

[0223] The weighted feature is calculated by the output gate to obtain output gate information:

[0224] o t = σ(W0[h t-1 ,x t ]+b0)

[0225] wherein o t represents the output gate information; W0 represents an output gate weight matrix; and b0 represents an output gate bias term.

[0226] According to the output gate information and the cell state of the current time point, a weighted feature decoded feature is calculated:

[0227] h t = o t tanh(C t )

[0228] wherein h t represents the weighted feature decoded feature; and h t is further subjected to a linear layer to obtain a predicted sequence of the trend sequence; and the cell state of the current time point is subjected to a tanh() function to normalize the output structure to between -1 and 1.

[0229] According to the system of the second aspect of the application, the fifth processing module 105 is specifically configured to apply a multi-layer fully connected neural network to predict the periodic sequence to obtain a predicted sequence of the periodic sequence:

[0230] S predicted = F3(S t )

[0231] wherein S predicted represents the predicted sequence of the periodic sequence; and F3() represents the multi-layer fully connected neural network.

[0232] The predicted sequence of the trend sequence, the predicted sequence of the periodic sequence, and the residual sequence are summed to obtain a predicted sequence of the water level of the post-pump station pool:

[0233] y predicted = T predicted +S predicted +R t

[0234] wherein y predicted represents the predicted sequence of the water level of the post-pump station pool; T predicted represents the predicted sequence of the trend sequence;

[0235] The model is updated in gradient according to the error between the predicted water level value and the actual water level, and the prediction accuracy of the neural network is improved:

[0236] E = å (y predicted -y true ) 2

[0237]

[0238] Wherein, y true represents the actual water level; E represents the sum of squares of errors; W new is the updated parameter, W old is the current parameter, and η is the learning rate, is the partial derivative of the error function with respect to the current parameter. The model can automatically adjust the parameters in the network according to the error between the prediction result and the actual result of each training sample.

[0239] The third aspect of the present application discloses an electronic device. The electronic device comprises a memory and a processor, and the memory stores a computer program. When the processor executes the computer program, the steps of any one of the pump station back pool water level prediction methods based on multi-dimensional sequence data correlation analysis disclosed in the first aspect of the present application are implemented.

[0240] Figure 3 As shown in FIG. 1, Figure 3 the electronic device comprises a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the electronic device is used for wired or wireless communication with external terminals. Wireless communication can be achieved through WIFI, operator network, near field communication (NFC) or other technologies. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen. The input device of the electronic device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the electronic device, or an external keyboard, touchpad or mouse, etc.

[0241] Those skilled in the art can understand that Figure 3 the structure shown in FIG. 1 is only a structural diagram of part of the technical solution of the present disclosure, and does not constitute a limitation on the electronic device to which the technical solution of the present application is applied. The specific electronic device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0242] The fourth aspect of the present application discloses a computer readable storage medium. The computer readable storage medium stores a computer program. The computer program is executed by a processor to implement the steps in the pump station back pool water level prediction method based on multi-dimensional sequence data correlation analysis in any one of the first aspect of the present application.

[0243] Please note that the technical features of the above embodiments can be combined in any way. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the description. The above embodiments only express several embodiments of the present application, which are described in detail and in detail, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for predicting water level in a pump station back pool based on multidimensional sequence data correlation analysis, characterized in that: The method comprises: Step S1: Collect pump station operation monitoring data and construct a pump station operation monitoring data set; based on the STL time series decomposition model, perform local weighted polynomial regression fitting on the pump station operation monitoring data set to decompose it into a trend sequence, a period sequence, and a residual sequence; Step S2: construct trend time series samples for the trend sequence by means of a sliding window; perform feature extraction on the trend time series samples using an encoder with three composite convolutional layers to obtain coding features; Step S3: using multiple parallel neural networks based on feature weighting mechanisms to extract multiple position weighted information of the coding features, and splicing the multiple position weighted information; multiplying the spliced ​​weighted information with the training parameter matrix to obtain weighted features; Step S4: decoding the weighted features based on a decoder of a recurrent neural network layer with a gating mechanism to obtain a prediction sequence of the trend sequence; Step S5: applying a multi-layer fully connected neural network to predict the periodic sequence to obtain a prediction sequence of the periodic sequence; summing the prediction sequence of the trend sequence, the prediction sequence of the periodic sequence, and the residual sequence to obtain a prediction sequence of the water level of the back pool of the pumping station; In step S1, the pump station operation monitoring data includes: Start-up time, voltage, current, water pump blade angle, pipeline pressure, water pump frequency, cumulative flow, reverse cumulative flow, active power, reactive power, unit vibration and swing, outlet pressure, water pump speed, inlet tank water level, front tank water level and rear tank water level; In step S2, the method of extracting features from the trend time series samples using an encoder with three composite convolutional layers to obtain coding features includes: Inputting the trend time series sample into a convolution layer with weight normalization of two first composite convolution layers connected in series to obtain a weight normalized convolution output; A residual structure is introduced, and a 1×1 convolution layer is used to adjust the dimension of the trend time series sample so that its dimension is the same as the dimension of the weighted normalized convolution output. After connecting the dimension-adjusted trend time series sample with the weighted normalized convolution output, an activation function is input to obtain the first composite convolution output. Inputting the first composite convolution output into a convolution layer with a hole connected in series with two second composite convolution layers to obtain a hole convolution output; After connecting the first composite convolution output and the dilated convolution output, an activation function is input to obtain a second composite convolution output; Inputting the second composite convolution output into a third composite convolution layer to obtain a third composite convolution output; the third composite convolution layer is a convolution layer with a hole consisting of two second composite convolution layers connected in series; After connecting the second composite convolution output and the third composite convolution output, the activation function is input to obtain the encoding feature.

2. The method for predicting water level in the back pool of a pumping station based on multidimensional sequence data correlation analysis according to claim 1, characterized in that: In step S2, the method of inputting the trend time series sample into a convolution layer with weight normalization of two first composite convolution layers connected in series to obtain a weight normalized convolution output includes: The structure of the convolutional layer with weight normalization of the first composite convolutional layer is: After ReLU function activation and Dropout random inactivation, nonlinearity is introduced and overfitting risk is reduced, and the convolution layer output with weight normalization of the first composite convolution layer is obtained after activation and inactivation. in, Represents the convolutional layer process output with weight normalization of the first composite convolutional layer; is the weight of the convolution kernel; b (1) is the bias term of the convolution kernel; c is the size of the convolution kernel; trend time series sample T τ =[T τ,1 ,T τ,2 ,...,T τ,n ]; Wherein, F1() represents the convolutional layer with weight normalization of the first composite convolutional layer; Represents the weight normalized convolution output.

3. The method for predicting water level in the back pool of a pumping station based on multidimensional sequence data correlation analysis according to claim 1, characterized in that: In step S2, the method of inputting the first composite convolution output into a convolution layer with a hole of two second composite convolution layers connected in series to obtain a void convolution output includes: The structure of the convolutional layer with a hole in the second composite convolutional layer is: After ReLU function activation and Dropout random inactivation, nonlinearity is introduced and the risk of overfitting is reduced, and the output of the second composite convolutional layer with a hole is obtained after activation and inactivation. in, represents the output of the convolutional layer process with holes in the second composite convolutional layer; f() represents the convolution kernel, d represents the dilation factor, τ-d·i represents the past direction, Indicates that the input data is convolved with a convolution kernel with a dilation factor of d; Wherein, F2() represents the convolutional layer with holes in the second composite convolutional layer; Represents the output of the dilated convolution.

4. The method for predicting water level in the back pool of a pumping station based on multidimensional sequence data correlation analysis according to claim 1, characterized in that: In step S3, the method of extracting the position weighted information of the coding feature using a neural network based on a feature weighting mechanism includes: Q=w Q Y t K=w K Y t V=w V Y t Among them, Q represents the query vector, K represents the key vector, and V represents the value vector; w Q 、w K and w V is the weight matrix, and the weight matrix of each neural network based on the feature weighting mechanism is different; Y t Represents the encoding feature; Through the query vector Q and key vector K, the attention level W of each position is obtained. in, is the transpose of K, d k Represents the dimension of the key vector K; softmax() is the normalized exponential function; A i (Q,K,V)=WV Among them, A i (Q, K, V) represents the position weighted information extracted by the i-th neural network based on the feature weighting mechanism.

5. The method for predicting water level in the back pool of a pumping station based on multidimensional sequence data correlation analysis according to claim 1, characterized in that: In step S4, the method for decoding the weighted features based on the decoder of the recurrent neural network layer with a gating mechanism includes: The weighted features are selected to forget information through the forget gate: f t =σ(W f [h t-1 ,x t ]+b f Among them, f t Indicates forgotten information; h t-1 Indicates the output at the previous moment; x t represents weighted features; σ(·) represents activation function, W f is the forget gate weight matrix, b f is the forget gate bias term; The weighted features are updated by the input gate: i t =σ(W i [h t-1 ,x t ]+b i ) Among them, i t Indicates the update amount; W i represents the input gate weight matrix; b i represents the input gate bias term; The weighted features are passed through the tanh layer to generate candidate values ​​for the current cell state; in, represents the candidate value of the unit state at the current moment; tanh() represents the tanh function; W c is connected to the previous moment h t-1 and the current time x t The weight matrix of b c Represents the bias term of the tanh layer; Calculate the current cell state based on the forgotten information, update amount, and candidate value vector: Among them, C t Indicates the unit state at the current moment; C t-1 Indicates the unit state at the previous moment; The weighted features are calculated through the output gate to calculate the output gate information: about t =σ(W0[h t-1 ,x t ]+b0) Among them, t Represents the output gate information; W0 represents the output gate weight matrix; b0 represents the output gate bias term; According to the output gate information and the current unit state, the weighted feature decoding feature is calculated: h t =o t fishy(C) t ) Among them, h t Represents the features after weighted feature decoding.

6. A pump station back pool water level prediction system based on multidimensional sequence data correlation analysis, characterized in that: The system adopts the method according to any one of claims 1 to 5, and the system includes: The first processing module is configured to collect pump station operation monitoring data and construct a pump station operation monitoring data set; based on the STL time series decomposition model, perform local weighted polynomial regression fitting on the pump station operation monitoring data set to decompose it into a trend series, a period series, and a residual series; The second processing module is configured to construct a trend time series sample for the trend sequence by means of a sliding window; perform feature extraction on the trend time series sample using an encoder with three composite convolutional layers to obtain coding features; A third processing module is configured to use multiple parallel neural networks based on a feature weighting mechanism to extract multiple position weighted information of the encoding feature, and splice the multiple position weighted information; multiply the spliced ​​weighted information by the training parameter matrix to obtain a weighted feature; a fourth processing module, configured to decode the weighted features based on a decoder of a recurrent neural network layer with a gating mechanism to obtain a prediction sequence of the trend sequence; The fifth processing module is configured to apply a multi-layer fully connected neural network to predict the periodic sequence to obtain a prediction sequence of the periodic sequence; sum the prediction sequence of the trend sequence, the prediction sequence of the periodic sequence and the residual sequence to obtain a prediction sequence of the water level of the rear pool of the pumping station.

7. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it implements the steps in the pump station back pool water level prediction method based on multidimensional sequence data correlation analysis as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the method for predicting the water level of the rear pool of a pumping station based on multidimensional sequence data correlation analysis according to any one of claims 1 to 5 are implemented.

Images