Performance parameter prediction method and system based on multilevel quantile recurrent neural network

Abstract

The invention relates to a performance parameter prediction method and system based on a multilevel quantile recurrent neural network, and belongs to the field of data processing. Comprising the steps that an equipment parameter table is acquired for data preprocessing to obtain target data, and the target data is a detailed table of time and prediction parameters; inputting target data into the trained multilevel quantile recurrent neural network model to obtain model output, placing a decoder with shared parameters in each loop layer in an encoder to create multi-step time sequence prediction, and calculating loss by adopting a quantile loss function, a global multi-layer perceptron and a local multi-layer perceptron are used for sensing at the decoder; and obtaining prediction parameters based on model output. According to the method, on the basis of the recurrent neural network, multi-step time sequence prediction and quantile regression are combined, a bifurcated sequence training strategy is used, accurate quantiles are predicted, the model is more stable, error accumulation of the recurrent prediction strategy is avoided, and the defect that the accumulated error of long-time prediction is too large is overcome.

Description

Technical Field

[0001] The present invention relates to a performance parameter prediction method and system based on a multi-level quantile recurrent neural network, belonging to the technical field of data processing. Background Art

[0002] A recursive neural network (RNN) is an artificial neural network (ANN) with a tree-like hierarchical structure in which network nodes recursively process input information according to the order in which they are connected. It is one of the deep learning algorithms.

[0003] At present, in the application process, a model has been developed that combines the direct prediction strategy (Multi-Horizon strategy) and quantile regression (Quantile-Regression, QR) based on the traditional recurrent neural network (RNN). Among them, the direct prediction strategy avoids the error accumulation of the recursive prediction strategy, and the use of quantile regression to predict specific quantiles of the distribution helps to make the optimal decision and can also quantify the uncertainty of the predicted value. At present, the multi-level quantile recurrent neural network (Multi-Horizon Quantile Recurrent Neural Net, MQ-RNN) prediction model has been successfully used in power load forecasting and electricity price forecasting problems, but has not yet been involved in the prediction of equipment performance parameters.

[0004] Therefore, there is an urgent need to design a performance parameter prediction method and system based on a multi-level quantile recurrent neural network that can solve the above technical problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a performance parameter prediction method and system based on a multi-level quantile recursive neural network. On the basis of traditional recursive neural networks, it combines multi-step time series prediction and quantile regression, uses a bifurcation sequence training strategy, predicts accurate quantiles, makes the model more stable, avoids the error accumulation of the recursive prediction strategy, and makes up for the deficiency of excessive cumulative error in long-term prediction.

[0006] The first aspect of the present invention provides a performance parameter prediction method based on a multi-level quantile recurrent neural network, which is special in that it includes the following steps:

[0007] Step 1: Obtain the equipment parameter table and perform data preprocessing to obtain target data, wherein the target data is a detailed table of time and prediction parameters;

[0008] Step 2: Input the target data into the trained multi-level quantile recurrent neural network model to obtain the model output.

[0009] Step 3: Place a decoder with shared parameters in each recurrent layer of the encoder to create a multi-step time series forecast, use the quantile loss function to calculate the loss, and use the global multi-layer perceptron and local multi-layer perceptron for perception in the decoder;

[0010] Step 4: Obtain prediction parameters of the performance to be tested of the current equipment based on the model output.

[0011] Preferably, the step 1 specifically includes:

[0012] Obtain the equipment parameter table to perform removal operations to obtain the time value t and the predicted parameter value y t ,in,

[0013] After removing the preset parameters to be measured in the equipment parameter table, x is obtained based on the values of all the remaining parameters. t .

[0014] Preferably, the training method of the multi-level quantile recurrent neural network model in step 2 comprises the following steps:

[0015] Acquiring historical data, wherein the historical data is row list data of time and performance parameters;

[0016] The historical data is grouped according to a preset ratio to obtain training data and verification data, wherein the time intervals of the training data and the verification data are different;

[0017] Training a multi-level quantile recurrent neural network initial model based on the training data and collecting multi-layer error gradients through back propagation;

[0018] When the number of training times reaches a preset number, the training ends, and after each training is completed, two verifications are performed using the verification data, where the preset number includes 500.

[0019] Preferably, the historical data is a row list data of time and performance parameters, specifically including:

[0020] Obtain a historical data table, and fuse the historical data table to obtain the row table data, wherein the fused row table data has N rows and M columns of data;

[0021] The fused dataset is Among them, i is the row number of the row list data;

[0022] N is the test value corresponding to the remaining M-1 performance parameters after removing the single target parameter value;

[0023] N is the target parameter value.

[0024] Preferably, the training of the multi-level quantile recurrent neural network initial model based on the training data and collecting multi-layer error gradients by back propagation specifically includes:

[0025] During training, the bifurcation sequence is represented as:

[0026] h t =encoder(x :t ,y :t );

[0027]

[0028] Among them, x :t Indicates the covariate with available time value "t", y :t The parameter value of the prediction parameter at time "t", h t is the hidden state value output by the encoder, Covariates available for "t:" time values, Indicates the quantile of the prediction corresponding to the prediction moment.

[0029] Preferably, step 3 places a decoder with shared parameters in each recurrent layer of the encoder to create a multi-step time series prediction, and uses a quantile loss function to calculate the loss, including the following steps:

[0030] The function to minimize the total quantile loss is:

[0031]

[0032] Among them, t represents the time value, "t+k" represents the predicted time, Q represents the quantiles predicted, and y t+k represents the known measured value at the prediction time, Indicates the quantile predicted at the prediction time, L q is the quantile loss function, which includes:

[0033]

[0034] Where y is a known measured value, is the predicted quantile, q is the set value, (·) + =max(0,·);

[0035] Preferably, step 3 uses a global multi-layer perceptron and a local multi-layer perceptron to perform perception at the decoder, comprising the following steps:

[0036] The input to the encoder is x :t and y :t , after encoding by the long short-term memory neural network, the hidden state value h is obtained t ;

[0037] The decoder uses a global multilayer perceptron and a local multilayer perceptron for decoding, where

[0038] The input of the global multilayer perceptron is h t and The output is c t+k and c a , where "t+k" represents the information obtained at the "t+k" time, c a To obtain global information;

[0039] The input of the local multilayer perceptron is c t+k 、c a and The output is the predicted value at each quantile at different time points in the future.

[0040] Preferably, the step 4 specifically includes:

[0041] The prediction expression is:

[0042]

[0043] Among them, y t+k represents the known measured value at the "t+k" prediction time, y :t The parameter value of the prediction parameter at time "t", x :t Indicates that the covariate with time value ":t" is available, Covariates available for time values "t:", x (s) It is a static invariant feature.

[0044] A second aspect of the present invention further provides a performance parameter prediction system based on a multi-level quantile recurrent neural network, comprising a memory and a processor, wherein the memory includes a performance parameter prediction method program based on a multi-level quantile recurrent neural network, and when the performance parameter prediction method program based on a multi-level quantile recurrent neural network is executed by the processor, the following steps are implemented:

[0045] Obtaining an equipment parameter table and performing data preprocessing to obtain target data, wherein the target data is a detailed table of time and prediction parameters;

[0046] The target data is input into the trained multi-level quantile recurrent neural network model to obtain the model output, wherein,

[0047] A decoder with shared parameters is placed in each recurrent layer of the encoder to create multi-step time series predictions. The loss is calculated using the quantile loss function, and the global multi-layer perceptron and local multi-layer perceptron are used for perception in the decoder.

[0048] Based on the model output, prediction parameters of the current equipment's performance to be tested are obtained.

[0049] The third aspect of the present invention provides a computer-readable storage medium, which includes a performance parameter prediction method program based on a multi-level quantile recurrent neural network for a machine. When the performance parameter prediction method program based on a multi-level quantile recurrent neural network is executed by a processor, the steps of a performance parameter prediction method based on a multi-level quantile recurrent neural network as described in any one of the above items are implemented.

[0050] The present invention discloses a performance parameter prediction method and system based on a multi-level quantile recursive neural network. On the basis of traditional recursive neural networks, it combines multi-step time series prediction and quantile regression, uses a bifurcation sequence training strategy to predict accurate quantiles, makes the model more stable, avoids the error accumulation of the recursive prediction strategy, and makes up for the deficiency of excessive cumulative error in long-term prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a performance parameter prediction method based on a multi-level quantile recurrent neural network according to the present invention;

[0052] Figure 2 A schematic diagram of three-dimensional prediction results of a performance parameter prediction method based on a multi-level quantile recurrent neural network according to the present invention;

[0053] Figure 3 A schematic diagram of two-dimensional prediction results of a performance parameter prediction method based on a multi-level quantile recurrent neural network according to the present invention;

[0054] Figure 4 This is a block diagram of a performance parameter prediction system based on a multi-level quantile recurrent neural network of the present invention. DETAILED DESCRIPTION

[0055] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0056] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0057] like Figure 1 As shown, this embodiment discloses a performance parameter prediction method based on a multi-level quantile recurrent neural network, comprising the following steps:

[0058] S102, obtaining an equipment parameter table and performing data preprocessing to obtain target data, wherein the target data is a detailed table of time and prediction parameters;

[0059] S104, inputting the target data into a trained multi-level quantile recurrent neural network model to obtain a model output;

[0060] S106, placing a decoder with shared parameters in each recurrent layer of the encoder to create a multi-step time series prediction, using a quantile loss function to calculate the loss, and using a global multi-layer perceptron and a local multi-layer perceptron for perception in the decoder;

[0061] S108: Obtain prediction parameters of the performance to be tested of the current equipment based on the model output.

[0062] It should be noted that, in this embodiment, the multi-level quantile recurrent neural network MQ-RNN is a prediction model that predicts future time series values through historical time series values and features. It mainly combines multi-step time series prediction and quantile regression on the traditional RNN, and uses a bifurcation sequence training strategy to predict accurate quantiles. Among them, MQ-RNN utilizes the expressiveness and temporal properties of sequence-to-sequence neural networks, the non-parametric properties of quantile regression and the efficiency of multi-step time series prediction to make the model more stable. Structurally, it is an encoder-decoder structure.

[0063] MQ-RNN uses a long short-term memory (LSTM) neural network to encode all historical information into hidden states. LSTM is a time-recursive neural network suitable for processing and predicting the states of events with relatively long intervals and delays in time series. As a special recursive neural network (RNN), LSTM can solve the problems of gradient vanishing and gradient exploding in traditional RNNs. In this embodiment, MQ-RNN designs two multilayer perceptrons (MLPs)—a global multilayer perceptron and a local multilayer perceptron—as decoders. MLP is a forward-structured artificial neural network that can be viewed as a directed graph. MLP consists of an input layer, a hidden layer, and an output layer. Each layer is fully connected to the next layer. Except for the input node, each node is a neuron with a nonlinear activation function. In the MQ-RNN network structure, the encoder at each time point is followed by a decoder.

[0064] Specifically, in this embodiment, the network model is trained and applied after the training is completed. Therefore, when applied, the equipment parameter table is obtained for data preprocessing to obtain target data, wherein the target data is a detailed table of time and prediction parameters, that is, a set of all parameters except the prediction parameters in the known time, and a set of prediction parameters corresponding to the future prediction time. The target data is then input into the trained multi-level quantile recursive neural network model to obtain the model output. Accordingly, the model processing process is generally consistent with the training, wherein a decoder with shared parameters is placed in each loop layer in the encoder to create a multi-step time series prediction, the quantile loss function is used to calculate the loss, and the global multi-layer perceptron and the local multi-layer perceptron are used for perception in the decoder. Finally, the prediction parameters of the current equipment's performance to be tested are obtained based on the model output.

[0065] According to an embodiment of the present invention, the equipment parameter table is obtained and data preprocessed to obtain target data, wherein the target data is a detailed table of time and prediction parameters, specifically including:

[0066] Obtain the equipment parameter table to perform removal operations to obtain the time value t and the predicted parameter value y t ,in,

[0067] After removing the preset parameters to be measured in the equipment parameter table, x is obtained based on the values of all the remaining parameters. t .

[0068] It should be noted that in this embodiment, when applied, data processing can be performed based on the equipment parameter table to remove the preset parameters to be measured in the equipment parameter table, and then x is obtained based on the values of all the remaining parameters. t , and get the time value t and the predicted parameter value y at the same timet .

[0069] According to an embodiment of the present invention, the training method of the multi-level quantile recurrent neural network model includes the following steps:

[0070] Acquiring historical data, wherein the historical data is row list data of time and performance parameters;

[0071] The historical data is grouped according to a preset ratio to obtain training data and verification data, wherein the time intervals of the training data and the verification data are different;

[0072] Training a multi-level quantile recurrent neural network initial model based on the training data and collecting multi-layer error gradients through back propagation;

[0073] When the number of training times reaches a preset number, the training ends, and after each training is completed, two verifications are performed using the verification data, where the preset number includes 500.

[0074] It should be noted that, in this embodiment, how to train the model is described, wherein, first, consistent with the application level, the data needs to be preprocessed, specifically, historical data is obtained, wherein the historical data is row list data of time and performance parameters, so that training data and verification data are obtained based on the historical data and grouped according to a preset ratio, wherein the time intervals of the training data and the verification data are different, for example, the ratio can be "4:1", and during training, the multi-level quantile recursive neural network initial model is trained based on the training data, and the multi-layer error gradient is collected through back propagation, and the training ends when the number of training times reaches the preset number, for example, the training is stopped after the training reaches "500" times, and the verification data is used to perform two verifications after each training is completed, wherein the number of layers of the network model is two, wherein the number of neurons in the encoder hidden layer is "50", and the number of neurons in the decoder hidden layer is "20", which can predict the values of various performance parameters in the next five years.

[0075] According to an embodiment of the present invention, the acquisition of historical data, wherein the historical data is a row list data of time and performance parameters, specifically includes:

[0076] Obtain a historical data table, and fuse the historical data table to obtain the row table data, wherein the fused row table data has N rows and M columns of data;

[0077] The fused dataset is Among them, i is the row number of the row list data;

[0078] N is the test values corresponding to the remaining M-1 performance parameters after removing the single target parameter value;

[0079] N is the target parameter value.

[0080] It should be noted that, in this embodiment, the historical data table includes a total of 23 data tables, which are merged into a record row list data with 102 rows and 308 columns, where N=102 and M=308. N=102 is the target parameter value, and after removing the single target parameter value, the remaining M-1=307 performance parameters corresponding to the test values are N=102.

[0081] Furthermore, the above embodiment illustrates that the ratio can be "4:1". Therefore, the time periods "1" to "82" in the fused row list are selected as the training set to fit the model, and the remaining "20" data are used for verification. That is, the data from 2006 to 2009 are used for training to verify the data from 2010 and 2011. The accuracy verification results of some equipment parameters are shown in Table 1.

[0082] Table 1. Detailed list of accuracy verification results of some equipment parameters

[0083]

[0084] According to an embodiment of the present invention, the training of the multi-level quantile recurrent neural network initial model based on the training data and collecting multi-layer error gradients through back propagation specifically includes:

[0085] During the training process, a bifurcation sequence is used for training, and the model is trained by minimizing the total quantile loss function, where the total quantile loss function is minimized as follows:

[0086]

[0087] Among them, t represents the time value, "t+k" represents the predicted time, Q represents the quantiles predicted, and y t+k represents the known measured value at the prediction time, Indicates the quantile predicted at the prediction time, L q is the quantile loss function, which includes:

[0088]

[0089] Where y is a known measured value, is the predicted quantile, q is the set value, (·) + =max(0,·);

[0090] During training, the bifurcation sequence is represented as:

[0091] h t =encoder(x :t ,y :t );

[0092]

[0093] Among them, x :t Indicates the covariate with available time value "t", y :t The parameter value of the prediction parameter at time "t", h t is the hidden state value output by the encoder, Covariates available for "t:" time values, Indicates the quantile of the prediction corresponding to the prediction moment.

[0094] It should be noted that, in this embodiment, the bifurcation sequence training method is applied during training, and a decoder with shared parameters is placed in each loop layer (time point) in the encoder to create a multi-step time series forecast, and the target loss of the response is calculated at the same time. Combined with the idea of quantile regression, the quantile loss function is adopted, and the model is trained by minimizing the quantile total loss function. The training data is input into the model for training, and the multi-layer error gradient is collected through back propagation to obtain the quantile predicted by the final prediction model.

[0095] According to an embodiment of the present invention, during the training process, the prediction expression is:

[0096]

[0097] Among them, y t+k represents the known measured value at the "t+k" prediction time, y :t The parameter value of the prediction parameter at time "t", x :t Indicates that the covariate with time value ":t" is available, Covariates available for time values "t:", x (s) It is a static and invariant feature;

[0098] The input to the encoder is x :t and y :t , after encoding by the long short-term memory neural network, the hidden state value h is obtained t ;

[0099] The decoder uses a global multilayer perceptron and a local multilayer perceptron for decoding, where

[0100] The input of the global multilayer perceptron is h t and The output is c t+k and c a, where "t+k" represents the information obtained at the "t+k" time, c a To obtain global information;

[0101] The input of the local multilayer perceptron is c t+k 、c a and The output is the predicted value at each quantile at different time points in the future.

[0102] It should be noted that in this embodiment, when the MQ-RNN prediction model is used to predict the performance parameters, the input of the encoder is and y :t , after LSTM encoding, the hidden state h is output t In the decoder, two multilayer perceptrons are used - a global multilayer perceptron and a local multilayer perceptron. The input of the global multilayer perceptron is the hidden state h t and The output is c t+K and c a , where c t+K represents the information obtained at the future time t+K,K=1,2,K,k, c a Used to obtain global information, the formula is as follows: Among them, the input of the local multilayer perceptron is c t+K 、c a and Output the predicted values at each quantile at different time points in the future, q (g) Represents each quantile in the quantile list, and the public representation is as follows:

[0103] Specifically, an example is given with reference to the parameters in Table 1. Figure 2 and Figure 3 As shown in the figure, the x-axis of the three-dimensional prediction result graph is the measurement time, the y-axis is the value of the target performance parameter, and the z-axis is the probability density of the predicted value. The trend of the parameter in the next five years is predicted, and the predicted value and probability density of the parameter in the next five years are given. Due to the large number of performance parameters, only the corresponding "6" typical performance parameter prediction results in Table 1 are selected for display. The prediction results are uniformly retained to four decimal places. The blue broken line in the figure is the predicted value with the largest probability density. Figure 2 As shown in Figure 2, as the prediction time increases, the variance of the prediction value based on the multi-level quantile recurrent neural network MQ-RNN method does not increase as much as the prediction result of the GPR (Gaussian Process Regression, GPR) model, so it is more suitable for long-term prediction; at the same time, Figure 3As shown in the figure, the interval prediction results of some performance parameters of the equipment. The two-dimensional prediction results can be used as a supplement to the three-dimensional prediction result diagram, and the predicted values of the parameters can be clearly displayed in the diagram.

[0104] Figure 4 A block diagram of a performance parameter prediction system based on a multi-level quantile recurrent neural network of the present invention is shown.

[0105] like Figure 4 As shown, the present invention discloses a performance parameter prediction system based on a multi-level quantile recurrent neural network, including a memory and a processor. The memory includes a performance parameter prediction method program based on a multi-level quantile recurrent neural network. When the performance parameter prediction method program based on a multi-level quantile recurrent neural network is executed by the processor, the following steps are implemented:

[0106] Obtaining an equipment parameter table and performing data preprocessing to obtain target data, wherein the target data is a detailed table of time and prediction parameters;

[0107] The target data is input into the trained multi-level quantile recurrent neural network model to obtain the model output, wherein,

[0108] A decoder with shared parameters is placed in each recurrent layer of the encoder to create multi-step time series predictions. The loss is calculated using the quantile loss function, and the global multi-layer perceptron and local multi-layer perceptron are used for perception in the decoder.

[0109] Based on the model output, prediction parameters of the current equipment's performance to be tested are obtained.

[0110] It should be noted that, in this embodiment, the multi-level quantile recurrent neural network MQ-RNN is a prediction model that predicts future time series values through historical time series values and features. It mainly combines multi-step time series prediction and quantile regression on the traditional RNN, and uses a bifurcation sequence training strategy to predict accurate quantiles. Among them, MQ-RNN utilizes the expressiveness and temporal properties of sequence-to-sequence neural networks, the non-parametric properties of quantile regression and the efficiency of multi-step time series prediction to make the model more stable. Structurally, it is an encoder-decoder structure.

[0111] MQ-RNN uses a long short-term memory (LSTM) neural network to encode all historical information into hidden states. LSTM is a time-recursive neural network suitable for processing and predicting the states of events with relatively long intervals and delays in time series. As a special recursive neural network (RNN), LSTM can solve the problems of gradient vanishing and gradient exploding in traditional RNNs. In this embodiment, MQ-RNN designs two multilayer perceptrons (MLPs)—a global multilayer perceptron and a local multilayer perceptron—as decoders. MLP is a forward-structured artificial neural network that can be viewed as a directed graph. MLP consists of an input layer, a hidden layer, and an output layer. Each layer is fully connected to the next layer. Except for the input node, each node is a neuron with a nonlinear activation function. In the MQ-RNN network structure, the encoder at each time point is followed by a decoder.

[0112] Specifically, in this embodiment, the network model is trained and applied after the training is completed. Therefore, when applied, the equipment parameter table is obtained for data preprocessing to obtain target data, wherein the target data is a detailed table of time and prediction parameters, that is, a set of all parameters except the prediction parameters in the known time, and a set of prediction parameters corresponding to the future prediction time. The target data is then input into the trained multi-level quantile recursive neural network model to obtain the model output. Accordingly, the model processing process is generally consistent with the training, wherein a decoder with shared parameters is placed in each loop layer in the encoder to create a multi-step time series prediction, the quantile loss function is used to calculate the loss, and the global multi-layer perceptron and the local multi-layer perceptron are used for perception in the decoder. Finally, the prediction parameters of the current equipment's performance to be tested are obtained based on the model output.

[0113] According to an embodiment of the present invention, the equipment parameter table is obtained and data preprocessed to obtain target data, wherein the target data is a detailed table of time and prediction parameters, specifically including:

[0114] Obtain the equipment parameter table to perform removal operations to obtain the time value t and the predicted parameter value y t ,in,

[0115] After removing the preset parameters to be measured in the equipment parameter table, x is obtained based on the values of all the remaining parameters. t .

[0116] It should be noted that in this embodiment, when applied, data processing can be performed based on the equipment parameter table to remove the preset parameters to be measured in the equipment parameter table, and then x is obtained based on the values of all the remaining parameters. t , and get the time value t and the predicted parameter value y at the same timet .

[0117] According to an embodiment of the present invention, the training method of the multi-level quantile recurrent neural network model includes the following steps:

[0118] Acquiring historical data, wherein the historical data is row list data of time and performance parameters;

[0119] The historical data is grouped according to a preset ratio to obtain training data and verification data, wherein the time intervals of the training data and the verification data are different;

[0120] Training a multi-level quantile recurrent neural network initial model based on the training data and collecting multi-layer error gradients through back propagation;

[0121] When the number of training times reaches a preset number, the training ends, and after each training is completed, two verifications are performed using the verification data, where the preset number includes 500.

[0122] It should be noted that, in this embodiment, how to train the model is described, wherein, first, consistent with the application level, the data needs to be preprocessed, specifically, historical data is obtained, wherein the historical data is row list data of time and performance parameters, so that training data and verification data are obtained based on the historical data and grouped according to a preset ratio, wherein the time intervals of the training data and the verification data are different, for example, the ratio can be "4:1", and during training, the multi-level quantile recursive neural network initial model is trained based on the training data, and the multi-layer error gradient is collected through back propagation, and the training ends when the number of training times reaches the preset number, for example, the training is stopped after the training reaches "500" times, and the verification data is used to perform two verifications after each training is completed, wherein the number of layers of the network model is two, wherein the number of neurons in the encoder hidden layer is "50", and the number of neurons in the decoder hidden layer is "20", which can predict the values of various performance parameters in the next five years.

[0123] According to an embodiment of the present invention, the acquisition of historical data, wherein the historical data is a row list data of time and performance parameters, specifically includes:

[0124] Obtain a historical data table, and fuse the historical data table to obtain the row table data, wherein the fused row table data has N rows and M columns of data;

[0125] The fused dataset is Among them, i is the row number of the row list data;

[0126] N is the test values corresponding to the remaining M-1 performance parameters after removing the single target parameter value;

[0127] N is the target parameter value.

[0128] It should be noted that, in this embodiment, the historical data table includes a total of 23 data tables, which are merged into a record row list data with 102 rows and 308 columns, where N=102 and M=308. N=102 is the target parameter value, and after removing the single target parameter value, the remaining M-1=307 performance parameters corresponding to the test values are N=102.

[0129] Furthermore, the above embodiment illustrates that the ratio can be "4:1". Therefore, the time periods "1" to "82" in the fused row list are selected as the training set to fit the model, and the remaining "20" data are used for verification. That is, the data from 2006 to 2009 are used for training to verify the data from 2010 and 2011. The accuracy verification results of some equipment parameters are shown in Table 1.

[0130] Table 1. Detailed list of accuracy verification results of some equipment parameters

[0131]

[0132]

[0133] According to an embodiment of the present invention, the training of the multi-level quantile recurrent neural network initial model based on the training data and collecting multi-layer error gradients through back propagation specifically includes:

[0134] During the training process, a bifurcation sequence is used for training, and the model is trained by minimizing the total quantile loss function, where the total quantile loss function is minimized as follows:

[0135]

[0136] Among them, t represents the time value, "t+k" represents the predicted time, Q represents the quantiles predicted, and y t+k represents the known measured value at the prediction time, Indicates the quantile predicted at the prediction time, L q is the quantile loss function, which includes:

[0137]

[0138] Where y is a known measured value, is the predicted quantile, q is the set value, (·) + =max(0,·);

[0139] During training, the bifurcation sequence is represented as:

[0140] h t =encoder(x :t ,y :t );

[0141]

[0142] Among them, x :t Indicates the covariate with available time value "t", y :t The parameter value of the prediction parameter at time "t", h t is the hidden state value output by the encoder, covariates available for time values of "t:", Indicates the quantile of the prediction corresponding to the prediction moment.

[0143] It should be noted that, in this embodiment, the bifurcation sequence training method is applied during training, and a decoder with shared parameters is placed in each loop layer (time point) in the encoder to create a multi-step time series forecast, and the target loss of the response is calculated at the same time. Combined with the idea of quantile regression, the quantile loss function is adopted, and the model is trained by minimizing the quantile total loss function. The training data is input into the model for training, and the multi-layer error gradient is collected through back propagation to obtain the quantile predicted by the final prediction model.

[0144] According to an embodiment of the present invention, during the training process, the prediction expression is:

[0145]

[0146] Among them, y t+k represents the known measured value at the "t+k" prediction time, y:t is the parameter value of the prediction parameter at "t" time, and x :t Indicates that the covariate with time value ":t" is available, Covariates available for time values "t:", x (s) It is a static and invariant feature;

[0147] The input to the encoder is x :t and y:t, and the hidden state value h is obtained after encoding by the long short-term memory neural network t ;

[0148] The decoder uses a global multilayer perceptron and a local multilayer perceptron for decoding, where

[0149] The input of the global multilayer perceptron is h t and The output is c t+k and ca , where "t+k" represents the information obtained at the "t+k" time, c a To obtain global information;

[0150] The input of the local multilayer perceptron is c t+k 、c a and The output is the predicted value at each quantile at different time points in the future.

[0151] It should be noted that in this embodiment, when the MQ-RNN prediction model is used to predict the performance parameters, the input of the encoder is and y :t , after LSTM encoding, the hidden state h is output t In the decoder, two multilayer perceptrons are used - a global multilayer perceptron and a local multilayer perceptron. The input of the global multilayer perceptron is the hidden state h t and The output is c t+K and c a , where c t+K represents the information obtained at the future time t+K,K=1,2,K,k, c a Used to obtain global information, the formula is as follows: Among them, the input of the local multilayer perceptron is c t+K 、c a and Output the predicted values at each quantile at different time points in the future, q (g) Represents each quantile in the quantile list, and the public representation is as follows:

[0152] Specifically, an example is given with reference to the parameters in Table 1. Figure 2 and Figure 3 As shown in the figure, the x-axis of the three-dimensional prediction result graph is the measurement time, the y-axis is the value of the target performance parameter, and the z-axis is the probability density of the predicted value. The trend of the parameter in the next five years is predicted, and the predicted value and probability density of the parameter in the next five years are given. Due to the large number of performance parameters, only the corresponding "6" typical performance parameter prediction results in Table 1 are selected for display. The prediction results are uniformly retained to four decimal places. The blue broken line in the figure is the predicted value with the largest probability density. Figure 2 As shown in Figure 2, as the prediction time increases, the variance of the prediction value based on the multi-level quantile recurrent neural network MQ-RNN method does not increase as much as the prediction result of the GPR (Gaussian Process Regression, GPR) model, so it is more suitable for long-term prediction; at the same time, Figure 3As shown in the figure, the interval prediction results of some performance parameters of the equipment. The two-dimensional prediction results can be used as a supplement to the three-dimensional prediction result diagram, and the predicted values of the parameters can be clearly displayed in the diagram.

[0153] The third aspect of the present invention provides a computer-readable storage medium, which includes a performance parameter prediction method program based on a multi-level quantile recurrent neural network. When the performance parameter prediction method program based on a multi-level quantile recurrent neural network is executed by a processor, the steps of a performance parameter prediction method based on a multi-level quantile recurrent neural network as described in any one of the above items are implemented.

[0154] The present invention discloses a performance parameter prediction method and system based on a multi-level quantile recursive neural network. On the basis of traditional recursive neural networks, it combines multi-step time series prediction and quantile regression, uses a bifurcation sequence training strategy to predict accurate quantiles, makes the model more stable, avoids the error accumulation of the recursive prediction strategy, and makes up for the deficiency of excessive cumulative error in long-term prediction.

[0155] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0156] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0157] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0158] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0159] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

Claims

1. A performance parameter prediction method based on a multi-level quantile recurrent neural network, characterized in that: The following steps are involved: Step 1: Obtain the equipment parameter table and perform data preprocessing to obtain target data, wherein the target data is a detailed table of time and prediction parameters; Step 2: Input the target data into the trained multi-level quantile recurrent neural network model to obtain the model output, where: Step 3: Place a decoder with shared parameters in each recurrent layer of the encoder to create a multi-step time series forecast, use the quantile loss function to calculate the loss, and use the global multi-layer perceptron and local multi-layer perceptron for perception in the decoder; Step 4: Obtain prediction parameters of the performance to be tested of the current equipment based on the model output.

2. A performance parameter prediction method based on a multi-level quantile recurrent neural network according to claim 1, characterized in that: The step 1 specifically includes: Obtain the equipment parameter table to perform removal operations to obtain the time value t and the predicted parameter value y t ,in, After removing the preset parameters to be measured in the equipment parameter table, x is obtained based on the values of all the remaining parameters. t .

3. The performance parameter prediction method based on multi-level quantile recurrent neural network according to claim 1 is characterized in that: The training method of the multi-level quantile recurrent neural network model in step 2 includes the following steps: Acquiring historical data, wherein the historical data is row list data of time and performance parameters; The historical data is grouped according to a preset ratio to obtain training data and verification data, wherein the time intervals of the training data and the verification data are different; Training a multi-level quantile recurrent neural network initial model based on the training data and collecting multi-layer error gradients through back propagation; When the number of training times reaches a preset number, the training ends, and after each training is completed, two verifications are performed using the verification data, where the preset number includes 500.

4. The performance parameter prediction method based on multi-level quantile recurrent neural network according to claim 3 is characterized in that: The historical data is obtained, wherein the historical data is a row list data of time and performance parameters, specifically including: Obtain a historical data table, and fuse the historical data table to obtain the row table data, wherein the fused row table data has N rows and M columns of data; The fused dataset is Among them, i is the row number of the row list data; are the test values corresponding to the remaining M-1 performance parameters after removing the single target parameter value; is the target parameter value.

5. The performance parameter prediction method based on multi-level quantile recurrent neural network according to claim 4 is characterized in that: The training of the multi-level quantile recurrent neural network initial model based on the training data and collecting multi-layer error gradients through back propagation specifically includes: During training, the bifurcation sequence is represented as: h t =encoder(x :t ,y :t ); Among them, x :t Indicates the covariate with available time value "t", y :t The parameter value of the prediction parameter at time "t", h t is the hidden state value output by the encoder, covariates available for time values of "t:", Indicates the quantile of the prediction corresponding to the prediction moment.

6. The performance parameter prediction method based on multi-level quantile recurrent neural network according to claim 5, characterized in that: The decoder with shared parameters is placed in each recurrent layer of the encoder to create multi-step time series predictions, and the loss is calculated using the quantile loss function: The function to minimize the total quantile loss is: Among them, t represents the time value, "t+k" represents the predicted time, Q represents the quantiles predicted, and y t+k represents the known measured value at the prediction time, Indicates the quantile predicted at the prediction time, L q is the quantile loss function, which includes: Where y is a known measured value, is the predicted quantile, q is the set value, (·) + =max(0,·).

7. The performance parameter prediction method based on multi-level quantile recurrent neural network according to claim 6, characterized in that: The decoder uses a global multi-layer perceptron and a local multi-layer perceptron for perception, specifically including: The input to the encoder is x :t and y :t , after encoding by the long short-term memory neural network, the hidden state value h is obtained t ; The decoder uses a global multilayer perceptron and a local multilayer perceptron for decoding, where The input of the global multilayer perceptron is h t and The output is c t+k and c a , where "t+k" represents the information obtained at the "t+k" time, c a To obtain global information; The input of the local multilayer perceptron is c t+k 、c a and The output is the predicted value at each quantile at different time points in the future.

8. The performance parameter prediction method based on multi-level quantile recurrent neural network according to claim 7, characterized in that: Based on the model output, the predicted parameters of the current equipment's performance to be tested are obtained, specifically including: Among them, y t+k represents the known measured value at the "t+k" prediction time, y :t The parameter value of the prediction parameter at time "t", x :t Indicates that the covariate with time value ":t" is available. Covariates available for time values "t:", x (s) It is a static invariant feature.

9. A performance parameter prediction system based on a multi-level quantile recurrent neural network, characterized in that: The system comprises a memory and a processor, wherein the memory comprises a performance parameter prediction method program based on a multi-level quantile recurrent neural network, and when the performance parameter prediction method program based on a multi-level quantile recurrent neural network is executed by the processor, the following steps are implemented: Obtaining an equipment parameter table and performing data preprocessing to obtain target data, wherein the target data is a detailed table of time and prediction parameters; The target data is input into the trained multi-level quantile recurrent neural network model to obtain the model output, wherein, A decoder with shared parameters is placed in each recurrent layer of the encoder to create multi-step time series predictions. The loss is calculated using the quantile loss function, and the global multi-layer perceptron and local multi-layer perceptron are used for perception in the decoder. Based on the model output, prediction parameters of the current equipment's performance to be tested are obtained.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a performance parameter prediction method program based on a multi-level quantile recurrent neural network. When the performance parameter prediction method program based on a multi-level quantile recurrent neural network is executed by a processor, the steps of a performance parameter prediction method based on a multi-level quantile recurrent neural network as described in any one of claims 1 to 8 are implemented.

Images