Coordinated optimization multi-step prediction method based on target correlation space-time coding-task decoding

Abstract

The invention discloses a coordinated optimization multi-step prediction method based on target correlation space-time coding-task decoding, and the method comprises the steps: constructing a space-time encoder formed by stacking a plurality of space attention enhanced time sequence convolution modules, and carrying out the complete adaptive noise set empirical mode decomposition of a target value historical sequence; constructing a parameter sharing mechanism, and training a multi-step prediction model formed by a space-time encoder and a task specific decoder by using a multi-task learning normal form; training input is preprocessed historical industrial process production data, and gradient conflicts among multiple tasks are reduced by using a gradient adjustment mechanism and a dynamic weight distribution method; and preprocessing real-time industrial process production data, and inputting the data into the trained prediction model to obtain a prediction result. According to the method, the relationship among different prediction tasks can be effectively coordinated, the overall model performance is optimized, and challenges encountered in long-term prediction are more effectively processed by dynamically adjusting weight distribution and gradient flow among multiple tasks.

Description

Technical Field

[0001] The present invention relates to the field of industrial process monitoring, and in particular to a coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding. Background Art

[0002] In the sintering process of modern steel industry, accurate prediction of sintering endpoint is crucial to improve production efficiency and reduce energy consumption. The main challenge of sintering endpoint prediction lies in processing high-dimensional, nonlinear and dynamically changing data generated in the process. Traditional prediction methods, such as experience-based judgment and simple statistical models, often cannot effectively handle these complexities.

[0003] In recent years, deep learning methods have made significant progress in industrial process modeling. These methods, such as autoencoder (AE), deep belief network (DBN), convolutional neural network (CNN), recurrent neural network (RNN), transformer, etc., have demonstrated their powerful data processing and pattern recognition capabilities. In the industrial field, more and more deep learning-based models have been proposed for process detection or prediction. However, existing methods are still facing challenges in effectively extracting features from multiple levels. For example, in terms of how to more effectively mine the historical information of the target variable to enhance the ability of the prediction model, traditional deep learning methods usually ignore the rich information contained in the historical target data, which leads to the potential of the prediction model in terms of accuracy not being fully utilized. Summary of the invention

[0004] In view of the deficiencies of the prior art, the present invention proposes a coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding.

[0005] The specific technical solutions are as follows:

[0006] A coordinated optimization multi-step prediction method based on target-related spatiotemporal encoding-task decoding includes the following steps:

[0007] S1: extracting historical industrial process production data within a period of time from the database of the sintering plant, performing data preprocessing on the data, and obtaining preprocessed historical industrial process production data;

[0008] S2: Construct a spatiotemporal encoder composed of multiple spatial attention enhanced temporal convolution modules stacked together. Each spatial attention enhanced temporal convolution module includes a causal dilation convolution module and a spatial attention module. The input x y The output of the latter is connected through a one-dimensional convolution residual; the causal expansion convolution stacking module has several layers, and the input data of each layer passes through the causal expansion convolution layer to obtain its temporal characteristics and expand the receptive field, and is activated by the nonlinear activation function layer after weight normalization, and then outputted through the discard layer, and finally the output feature map O of time step t is obtained. y ;O y Input the spatial attention module, after the global attention layer extracts the global information of each channel, input it into the first fully connected layer, activated by the PReLU function, and the result is input into the second fully connected layer, activated by the sigmoid function, and outputs the output feature map z after spatial attention adjustment y ;

[0009] S3: Perform fully adaptive noise set empirical mode decomposition on the historical sequence of target values ​​to obtain multiple intrinsic mode functions and a residual component; input the intrinsic mode functions with variance contribution ratio greater than the threshold into the gated neural network to extract time series features and generate a hidden state for each time point; input the hidden state into the fully connected layer and convert it into a dimension and format suitable for task-specific decoder processing;

[0010] S4: Build a task-specific decoder and use the temporal attention mechanism to adaptively extract relevant information from the entire spatiotemporal encoder output in each prediction step, capturing hidden states that are more closely related to the prediction target; use memory units with independent weights in the final output layer, and each prediction step has an independent set of parameters; and finally output multi-step prediction results;

[0011] S5: The spatiotemporal encoder and task-specific decoder form a multi-step prediction model, build a parameter sharing mechanism, and use the multi-task learning paradigm to train the multi-step prediction model; the training input is the data obtained in S1, and the gradient adjustment mechanism and dynamic weight allocation method are used during training to reduce the gradient conflict between multiple tasks and obtain a trained multi-step prediction model;

[0012] S6: Obtain real-time industrial process production data, pre-process it using the method of S1, and input it into the trained multi-step prediction model to obtain the prediction results.

[0013] Furthermore, the S1 is specifically implemented through the following sub-steps:

[0014] (1.1) Acquiring production data within a period of time from a time series database of a sintering plant as raw data, wherein the production data includes material variables and process variables;

[0015] (1.2) Use the isolation forest method to detect outliers in the original data. For abnormal data, execute step (1.3); for the remaining data, jump to step (1.4);

[0016] (1.3) Use local linear regression method to process abnormal data to maintain the continuity of time series data;

[0017] (1.4) Integrate the normal data and the data obtained in step (1.3), and use wavelet decomposition denoising technology to filter out the noise;

[0018] (1.5) Apply the sliding window method to the time series data obtained in step (1.4) to generate a set of time series data with a specific time step and prediction range, that is, the preprocessed historical production data.

[0019] Furthermore, in S2, the causal expansion convolution layer obtains the temporal feature F(t) in the input data through causal convolution, which is expressed as:

[0020]

[0021] In the formula, w represents the weight vector, x represents the input data, and w i represents the i-th component of w, x t-d*i represents the input data of the td*ith time step; k represents the convolution kernel size, t represents the current time step; d represents the expansion coefficient;

[0022] The calculation expression of the weight normalization layer is as follows:

[0023]

[0024] Where W′ represents the standardized weight, g represents the learnable scaling parameter, and W d represents the original dilated convolution weight, ||W d || represents the L2 norm of the original dilated convolution weights;

[0025] When the causal expansion convolution stack module has two layers, its input-output relationship expression is as follows:

[0026]

[0027] In the formula, represents the output feature map of the causal expansion stacking module at time step t, represents the input feature vector of the causal expansion stacking module at time step t, C is the number of channels, and T is the time step; * represents the causal expansion convolution operation; represents the normalized causal dilated convolution weight matrix of the first causal dilated convolution layer, represents the normalized causal dilated convolution weight matrix of the second causal dilated convolution layer; PReLU(·) represents the parameterized linear unit activation function; m1 and m2 are two randomly generated mask vectors used to implement the dropout operation; ⊙ represents the Hadamard product, which is used to apply the mask vector to the output of the causal dilated convolution layer.

[0028] Furthermore, in S2, the input of the spatial attention module is the output of the causal expansion convolution stack module. t , O t After the global information of each channel is extracted by the global attention layer, it is input into the first fully connected layer and activated by the PReLU function. The result is input into the second fully connected layer and activated by the sigmoid function. The output feature map after spatial attention adjustment is output. The expression is as follows:

[0029]

[0030] z t =O t ⊙σ(W2δ(W1I));

[0031] In the formula, represents the global information of each channel c, c = 1, 2, 3, ..., C, C is the number of channels; O t,c represents the output of the causal expansion convolution stack module at channel c; T is the time step, I represents a vector with C dimensions, and the value corresponding to each dimension is I c ; W1 represents the weight matrix of the first fully connected layer, W2 represents the weight matrix of the second fully connected layer; σ represents the sigmoid activation function.

[0032] Furthermore, in S3, the expression for performing fully adaptive noise set empirical mode decomposition on the target value history sequence is as follows:

[0033]

[0034] In the formula, y represents the historical sequence of target values, represents the pth intrinsic mode function, M is the total number of intrinsic mode functions obtained by the fully adaptive noise ensemble empirical mode decomposition, and R(t) represents the residual component;

[0035] The expression of the variance contribution ratio is as follows:

[0036]

[0037] In the formula, represents the variance of the pth intrinsic mode function, and Var(y) represents the total variance of the target sequence y.

[0038] Furthermore, in S3, the final hidden state ψ output by the gated neural network Q The input is a fully connected layer, which is linearly transformed and a nonlinear activation function is added to generate a new feature representation; the output of the fully connected layer is used as the initial state of the task-specific decoder, which includes an initial hidden state and initial cell state The intrinsic mode function of the input gated neural network is embedded in each gated neural network unit. and In this approach, the intrinsic modality is transformed into a dimension and format suitable for processing by a task-specific decoder.

[0039] Further, the S4 is specifically as follows:

[0040] The context feature matrix S output by the spatiotemporal encoder is t , divided into T context vectors s from the time dimension t-T+1 ,s t-T+2 , ..., s t , these context vectors are input one-to-one into each time step of the task-specific decoder, where corresponds to the input to the task-specific decoder s t-T+k In time step;

[0041] The temporal attention value of the hidden state of each task-specific decoder at time step t is expressed as follows:

[0042]

[0043]

[0044] In the formula, represents the temporal attention value of the k-th context vector at time step t, is the overall weight matrix, tanh represents the hyperbolic tangent activation function; c t-1 The weight matrix, c t-1 represents the hidden state of the task-specific decoder at time step t-1; For t-T+k The weight matrix, s t-T+k represents the output of the spatiotemporal encoder at time step t-T+k; is the bias term; express The weight value after softmax normalization at the same time step t, N represents the maximum prediction step length;

[0045] The input θ for each prediction step tIt is calculated by weighing the state variables of the corresponding step and the output of all time steps of the spatiotemporal encoder part, and is combined with the predicted label y of the previous time step before being input into the memory unit. t-1 Splicing, the expression is as follows:

[0046]

[0047] θ t =concatenate(u t ,y t-1 );

[0048] The independent weight memory units in the task-specific decoder generate the final multi-step prediction results as follows:

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056] In the formula, represents the forget gate of the mth memory unit at time step t, The weight matrix representing the input data of the forget gate, The weight matrix representing the hidden state of the previous memory unit of the forget gate, Represents the bias term of the forget gate; represents the input gate of the mth memory unit at time step t, The weight matrix representing the input data of the input gate, The weight matrix representing the hidden state of the previous memory cell of the input gate, represents the bias term of the input gate; represents the candidate cell state of the mth memory unit at time step t, The weight matrix representing the input data of the candidate cell states, The weight matrix representing the previous hidden state of the candidate cell state, A bias term representing the candidate cell state; represents the cell state of the mth memory unit at time step t, represents the cell state of the mth memory unit at time step t-1; o m t represents the output gate of the mth memory unit at time step t, The weight matrix representing the input data of the output gate, The weight matrix representing the hidden state of the previous memory unit of the output gate, represents the bias term of the output gate; h t represents the hidden state of the last memory unit at time step t, c t represents the cell state of the last memory unit of the decoder at time step t; represents the prediction result of the last memory unit at time step t; FC represents the fully connected layer, and each memory unit outputs h t After the fully connected layer, the predicted value corresponding to the prediction step length is finally obtained

[0057] Furthermore, in S5, the input data passes through the shared parameter layer and the task layer in sequence, the multi-step prediction model learns common features through parameter sharing in the shared parameter layer, and predicts the result of each step through an independent memory unit in the task layer, while maintaining the connection between different prediction steps through the flow of time series state information;

[0058] The gradient adjustment mechanism is as follows: Set the loss function of the n-th step prediction to M n (θ), loop to calculate the gradient of the prediction of the nth step The gradient of all other prediction steps In chronological order and Compare and determine whether the inner product of the two gradient vectors is less than 0. If not, there is no gradient conflict between the two gradient vectors and no adjustment is required; if so, there is a gradient conflict between the two gradient vectors and the calculation is exist projection on and from This projection is subtracted from the gradient to adjust

[0059] The dynamic weight allocation method is specifically as follows: set the number of training times to E, calculate the normalized weight of the n-th step prediction in all time step tasks in each round of training, n = 1, 2, 3, ..., N, N is the time step, and the expression is as follows:

[0060]

[0061]

[0062] Where M n (e-1) represents the prediction loss of the prediction of the nth step in the e-1th round, e = 1, 2, 3, ..., E; Mn (e-2) represents the prediction loss of the prediction of the nth step in the e-2th round, w n (e-1) represents the ratio of the two; λ n (e) is the normalized weight of the n-th step prediction in all time-step tasks during the e-th round of training; τ is a hyperparameter;

[0063] The loss function M(e) of the multi-step prediction model in the e-th round of training is:

[0064]

[0065] Where L is the total number of samples in the preprocessed historical industrial process production data; y n,l represents the true value of the lth sample in the nth step prediction task, Represents the predicted value of the lth sample in the nth step prediction task.

[0066] A coordinated optimized multi-step prediction device based on target-related spatiotemporal coding-task decoding comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the coordinated optimized multi-step prediction method based on target-related spatiotemporal coding-task decoding.

[0067] A computer-readable storage medium stores a program, which, when executed by a processor, implements the coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding.

[0068] The beneficial effects of the present invention are:

[0069] The present invention obtains the temporal features of historical data through causal expansion convolution, and captures the correlation of temporal features in a long time range; the spatial attention mechanism is used to more accurately capture and emphasize the feature information that contributes more to the prediction target, and residual connections are added to solve the gradient vanishing problem of deep networks; multiple spatial attention-enhanced temporal convolution layers are stacked through the spatiotemporal encoder to capture dependencies on different time scales from short-term to long-term; the output of the temporal module of the integrated fully adaptive noise set empirical mode decomposition is used as the starting state of the task-specific decoder to deeply mine the implicit information in the historical data of the target variable; a parameter sharing mechanism is constructed, and a multi-step prediction model composed of a spatiotemporal encoder and a task-specific decoder is trained using the paradigm of multi-task learning; a gradient adjustment mechanism and a dynamic weight allocation method are used to reduce gradient conflicts between multiple tasks; a temporal attention mechanism is used to adaptively extract the relevant information output by the encoder in each prediction step, thereby capturing the hidden state that is more closely related to the prediction target and outputting a more accurate multi-step prediction result. The present invention can effectively coordinate the relationship between different prediction tasks while maintaining the specific performance of each task, and optimize the overall model performance. By dynamically adjusting the weight distribution and gradient flow between multiple tasks, the model can more effectively handle the challenges encountered in long-term prediction, such as optimization conflicts and information loss between tasks, improve the model's adaptability to long-term prediction tasks, and enhance the model's flexibility in handling multi-task prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flow chart of a coordinated optimized multi-step prediction method based on target-related spatiotemporal coding-task decoding in an embodiment of the present invention.

[0071] Figure 2 4 is a flow chart of data preprocessing in an embodiment of the present invention.

[0072] Figure 3 It is a structural diagram of a temporal convolution module for spatial attention enhancement in an embodiment of the present invention.

[0073] Figure 4 It is a flow chart of extracting relevant information in a historical target sequence based on fully adaptive noise set empirical mode decomposition and gated neural network in an embodiment of the present invention.

[0074] Figure 5 Schematic diagram of a parameter sharing mechanism in an embodiment of the present invention.

[0075] Figure 6 Schematic diagram of gradient adjustment in an embodiment of the present invention.

[0076] Figure 7 It is a structural schematic diagram of a coordinated optimized multi-step prediction device based on target-related spatiotemporal coding-task decoding in an embodiment of the present invention. DETAILED DESCRIPTION

[0077] The present invention will be described in detail below according to the accompanying drawings and preferred embodiments, and the purpose and effect of the present invention will become more clear. The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0078] like Figure 1 As shown, a coordinated optimization multi-step prediction method based on target-related spatiotemporal encoding-task decoding includes the following steps:

[0079] S1: Extract historical industrial process production data within a period of time from the database of the sintering plant, perform data preprocessing on these data, and obtain preprocessed historical production data. Figure 2 The specific implementation process is as follows:

[0080] (1.1) Sensors collect production data during the sintering process and transmit them to the time series database through a programmable logic controller. The production data within a period of time is obtained from the time series database of the sintering plant as the original data. The production data includes material variables (such as iron ore mixing ratio, quicklime ratio, etc.) and process variables (such as ignition temperature, material layer thickness, etc.).

[0081] (1.2) Use the isolation forest method to detect outliers in the original data to identify and isolate abnormal data that may affect the performance of the model. For abnormal data, execute step (1.3); for the rest of the data, jump to step (1.4) for execution.

[0082] (1.3) The local linear regression method is used to process abnormal data to maintain the continuity of time series data.

[0083] (1.4) Integrate the normal data and the data obtained in step (1.3) and use wavelet decomposition denoising technology to filter out the noise in the data. This step helps to preserve the main structure of the data while minimizing the impact of random noise.

[0084] (1.5) Apply the sliding window method to the time series data obtained in step (1.4) to generate a set of time series data with a specific time step and prediction range, i.e., the preprocessed historical production data, which is used for the subsequent training, verification and testing of the prediction model.

[0085] S2: Construct a spatiotemporal encoder. The spatiotemporal encoder is obtained by stacking multiple temporal convolutional modules with spatial attention enhancement, so that the model can capture dependencies on different time scales from short-term to long-term, and improve the prediction performance of sequence tasks. In this embodiment, there are six temporal convolutional modules with spatial attention enhancement.

[0086] like Figure 3 As shown in the figure, the temporal convolution module with spatial attention enhancement includes: a causal expansion convolution stacking module and a spatial attention module, whose input and output are connected by a one-dimensional convolution residual. The temporal convolution module with spatial attention enhancement can more accurately capture and emphasize feature information that contributes more to the prediction target through the spatial attention mechanism, and solve the gradient vanishing problem of deep networks through residual connections. The architecture of the temporal convolution module with spatial attention enhancement is described in three parts below.

[0087] (1) The causal expansion convolution stack module has several layers (two layers in this embodiment), each of which includes: a causal expansion convolution layer, a weight normalization layer, a nonlinear activation function layer (specifically the PReLU function in this embodiment) and a dropout layer.

[0088] The causal dilated convolution layer obtains the temporal features of the preprocessed historical production data through causal convolution, expands the receptive field through dilated convolution, and captures the correlation of temporal features over a longer time range. By combining dilated convolution with causal convolution, the causal dilated convolution formed can cover longer input sequences without increasing parameters.

[0089] Specifically, the temporal feature F(t) in the preprocessed data is obtained through causal convolution, and its expression is as follows:

[0090]

[0091] In the formula, w represents the weight vector, x represents the input time series data (preprocessed data), and w i represents the i-th component of w, x t-d*i represents the input data of the td*ith time step; k represents the size of the convolution kernel, t represents the current time step; d represents the dilation coefficient, which determines the distance between adjacent elements in the convolution kernel. In the dilated convolution, this distance expands exponentially with the increase in the number of layers.

[0092] The calculation expression of the weight normalization layer is as follows:

[0093]

[0094] Where W′ represents the standardized weight, g represents the learnable scaling parameter, and W d represents the original dilated convolution weight, ‖W d ‖ represents the L2 norm of the original dilated convolution weights.

[0095] The input-output relationship expression of the two-layer stacked causal expansion convolution stack module is as follows:

[0096]

[0097] In the formula, represents the output feature map of the causal expansion stacking module at time step t, represents the input feature vector of the causal expansion stacking module at time step t, C is the number of channels, and T is the time step; * represents the causal expansion convolution operation; represents the normalized causal expansion convolution weight matrix of the first causal expansion convolution layer, represents the normalized causal dilated convolution weight matrix of the second causal dilated convolution layer; PReLU(·) represents the parameterized linear unit activation function; m1 and m2 are two randomly generated mask vectors used to implement the dropout operation; ⊙ represents the Hadamard product, which is used to apply the mask vector to the output of the causal dilated convolution layer.

[0098] (2) The spatial attention module includes: global attention layer, fully connected layer, PReLU activation function and sigmoid activation function. The output of the causal expansion convolution stack module is t , after the global attention layer extracts the global information of each channel, it is input into the first fully connected layer and activated by the PReLU function. The result is input into the second fully connected layer and activated by the sigmoid function. The output feature map after spatial attention adjustment is output. Its expression is as follows:

[0099]

[0100] z t =o t ⊙σ(W2PReLU(W1I))

[0101] In the formula, Represents the global information of each channel c, c = 1, 2, 3, ..., C; O t,c represents the output of the causal expansion convolution stack module at channel c; represents the output feature map after spatial attention adjustment, I represents a vector with C dimensions, and the value corresponding to each dimension is I c ; W1 represents the weight matrix of the first fully connected layer, W2 represents the weight matrix of the second fully connected layer; σ represents the sigmoid activation function.

[0102] (3) The expression of the convolution residual connection between the input and output of the temporal convolution module with spatial attention enhancement is as follows:

[0103]

[0104]

[0105] In the formula, represents the output feature map after 1×1 convolution, W 1*1 represents the weight matrix of 1×1 convolution, represents the bias vector of 1×1 convolution, Represents the final output feature map of the temporal convolution module with spatial attention enhancement, that is, the output of the temporal convolution module with spatial attention enhancement for all C channels at time step t.

[0106] When six spatial attention-enhanced temporal convolutional modules are stacked to form a spatiotemporal encoder, a single spatial attention-enhanced temporal convolutional module uses a 1D kernel to perform convolution operations in the time dimension, and the number of channels increases from 12 to 40, 80, and 160 to capture finer-grained features, and then decreases to 160, 80, and 40 to form the encoder output. This multi-scale design enables the spatiotemporal encoder to capture dependencies at different time scales from short-term to long-term, improving the prediction performance of sequence tasks.

[0107] Define a set representing different number of channels in Represents the number of channels of each spatial attention-enhanced temporal convolution module in the spatiotemporal encoder. The output of the entire spatiotemporal encoder can be expressed as:

[0108]

[0109] In the formula, represents the stacking operation, which is used to stack the output of each spatial attention-enhanced temporal convolution module along the channel dimension to form the final output of the spatiotemporal encoder; represents the output of the temporal convolutional module with spatial attention enhancement of c channels at time step t; represents the output of the spatiotemporal encoder at time step t, C ′ is the final number of channels output by the spatiotemporal encoder.

[0110] Through this structure, the spatiotemporal encoder effectively integrates long-term dependencies and spatial attention mechanisms, providing the prediction model with rich and effective spatiotemporal feature representations, thereby significantly improving the performance of multi-step prediction tasks.

[0111] S3: Perform fully adaptive noise ensemble empirical mode decomposition on the target value history sequence y to obtain M intrinsic mode functions and one residual component, which is expressed as follows:

[0112]

[0113] In the formula, represents the pth intrinsic mode function, and R(t) represents the residual component.

[0114] The more critical intrinsic mode functions are screened out according to the threshold, and the gated neural network is used to extract the time series features from them. The specific implementation process is as follows:

[0115] like Figure 6 As shown, for each intrinsic mode function, calculate its variance contribution ratio:

[0116]

[0117] In the formula, represents the variance of the pth intrinsic mode function, and Var(y) represents the total variance of the target sequence y.

[0118] The calculation of the variance contribution ratio provides a method to quantify the contribution of each intrinsic mode function to the overall target sequence variation. The reason for choosing the variance contribution ratio as the screening criterion is that it can effectively identify the intrinsic mode function that has the greatest influence on the prediction task. Specifically, by setting a predetermined threshold ∈ to filter the intrinsic mode function, only when the variance contribution ratio of the intrinsic mode function exceeds this threshold will it be retained for subsequent predictive modeling, and the remaining intrinsic modes will be screened out. The importance of this screening process is that it helps the model focus on those components that are most critical to the prediction, thereby improving the accuracy and efficiency of the prediction. The expression for this screening is as follows:

[0119] v=|{p:ρ p >∈}|

[0120] The important intrinsic mode functions obtained by screening are input into the gated neural network to extract the key time features. The gated neural network is a special type of recurrent neural network, which is particularly suitable for processing time series data because it can effectively capture the long-term dependencies in the sequence data.

[0121] Specifically, the gated neural network generates a hidden state for each time point through a series of computational steps. These hidden states capture the important features and patterns of the input sequence at different time points. For the prediction task, the last Q-th hidden state of the gated neural network is denoted by ψ Q , which is the final hidden state obtained after processing the entire intrinsic mode function sequence. The following are the calculation steps of the gated neural network:

[0122] z q =σ(W z ·[ψ q-1 ,IMF q ]+b z )

[0123] r q =σ(W r ·[ψ q-1 ,IMFq ]+b r )

[0124]

[0125]

[0126] In the formula, z q represents the update gate of the qth memory unit, which determines how much past information is retained in the hidden state; r q represents the reset gate of the qth memory unit. This parameter affects the amount of historical information retained; Represents the candidate hidden state of the qth memory unit, which contains the current input IMF q and information about past hidden states; ψ q represents the final hidden state of the qth memory unit, which combines historical information and current input; W z represents the weight matrix of the update gate, W r represents the weight matrix of the reset gate, W h represents the weight matrix for calculating the hidden state, b z represents the bias term of the update gate, b r represents the bias term of the reset gate, b h Represents the bias term for calculating the hidden state.

[0127] After processing the entire intrinsic mode function sequence, the final hidden state ψ output by the gated neural network is Q It contains comprehensive information of all important intrinsic mode functions, providing key contextual information for subsequent prediction steps. This hidden state contains not only information of the current time step, but also information of all previous time steps, thus capturing the dynamic changes of historical data.

[0128] The final hidden state ψ of the gated neural network Q The input is a fully connected layer that converts the information contained in it into a dimension and format suitable for processing by the task-specific decoder. The specific implementation process is as follows:

[0129] The fully connected layer pair ψ Q Perform linear changes and add nonlinear activation functions to generate new feature representations. The output of the fully connected layer is used as the initial state of the task-specific decoder, which includes the initial hidden state and initial cell state The conversion process can be expressed by the following formula:

[0130]

[0131] In the formula, FC represents the fully connected layer function, and ψ represents the sequence from the target value history to the initial hidden state. and initial cell state The function mapping includes the above-mentioned fully adaptive noise set empirical mode decomposition, mode function screening, gated neural network processing, and fully connected layer linear transformation operations. t Represents the tth value in the historical sequence of target values.

[0132] At the first time step t = t0 = 0, these key intrinsic modal functions are embedded into the initial hidden states of the memory cells of the task-specific decoder and cell status In the subsequent time step t ≥ t0, decoding is performed according to the method of S4.

[0133] S4: Build a task-specific decoder and use the temporal attention mechanism to adaptively extract relevant information from the entire spatiotemporal encoder output in each prediction step, thereby capturing hidden states that are more closely related to the prediction target and outputting more accurate multi-step prediction results. The specific implementation process is as follows:

[0134] A temporal attention mechanism is used to adaptively extract relevant information of the entire spatiotemporal encoder output in each prediction step, thereby capturing hidden states that are more closely related to the prediction target. Memory units with independent weights are used in the final output layer, and each prediction step has an independent set of parameters, so that the prediction model can more effectively learn and adapt to the specific features of each prediction step.

[0135] Using M t represents the context feature matrix from the spatiotemporal encoder (i.e., the output of the last temporal convolutional module), M t It can be divided into T context vectors s from the time dimension t-T+1 ,s t-T+2 , ..., s t , these context vectors are input one-to-one into each time step of the task-specific decoder, where corresponds to the input to the task-specific decoder s t-T+k In order to effectively capture the temporal patterns associated with each prediction step, we first compute the temporal attention value of the hidden state of each task-specific decoder at time step t:

[0136]

[0137]

[0138] Where N represents the maximum prediction step size, represents the temporal attention value of the k-th context vector at time step t, is the overall weight matrix; c t-1 The weight matrix, c t-1 represents the hidden state of the task-specific decoder at time step t-1; For t-T+k The weight matrix, s t-T+k represents the output of the spatiotemporal encoder at time step t-T+k; is the bias term, tanh represents the hyperbolic tangent activation function; express The weight value after softmax normalization at the same time step t.

[0139] The input θ for each prediction step t It is calculated by weighing the state variables of the corresponding step and the output of all time steps of the spatiotemporal encoder part, and is combined with the predicted label y of the previous time step before being input into the memory unit. t-1 Splicing, the expression is as follows:

[0140]

[0141] θ t =concatenate(u t ,y t-1 )

[0142] The independent weight memory units in the task-specific decoder generate the final multi-step prediction results as follows:

[0143]

[0144]

[0145]

[0146]

[0147]

[0148]

[0149]

[0150] In the formula, represents the forget gate of the mth memory unit at time step t, The weight matrix representing the input data of the forget gate, The weight matrix representing the hidden state of the previous memory unit of the forget gate, Represents the bias term of the forget gate; represents the input gate of the mth memory unit at time step t, The weight matrix representing the input data of the input gate, The weight matrix representing the hidden state of the previous memory cell of the input gate, represents the bias term of the input gate; represents the candidate cell state of the mth memory unit at time step t, The weight matrix representing the input data of the candidate cell states, The weight matrix representing the previous hidden state of the candidate cell state, A bias term representing the candidate cell state; represents the cell state of the mth memory unit at time step t, represents the cell state of the mth memory unit at time step t-1; o m t represents the output gate of the mth memory unit at time step t, The weight matrix representing the input data of the output gate, The weight matrix representing the hidden state of the previous memory unit of the output gate, represents the bias term of the output gate; h t represents the hidden state of the last memory unit at time step t, c t represents the cell state of the last memory unit of the decoder at time step t; represents the prediction result of the last memory unit at time step t; FC represents the fully connected layer, and each memory unit outputs h t After the fully connected layer, the predicted value corresponding to the prediction step length is finally obtained

[0151] S5: The spatiotemporal encoder and task-specific decoder form a multi-step prediction model, build a parameter sharing mechanism, and use the multi-task learning paradigm to train the multi-step prediction model. The specific implementation process is as follows:

[0152] The multi-step prediction framework based on parameter sharing mechanism is as follows Figure 5 As shown in the figure, the input data passes through the shared parameter layer and the task layer in sequence. The multi-step prediction model learns common features through parameter sharing in the shared parameter layer, predicts the results of each step through independent memory units in the task layer, and maintains the connection between different prediction steps through the flow of time series state information. Such a structure allows different tasks to share underlying features, thereby reducing the number of model parameters and the risk of overfitting.

[0153] The output of the rth task The calculation formula for (i.e. the prediction result of each task) is as follows:

[0154] Hshared =f(X;θ shared )

[0155]

[0156] Where X is the total input of the multi-step prediction model, H shared is the output after all shared parameter layers, θ shared are the parameters of the shared parameter layer, is the prediction result of the rth task, is the rth task-specific parameter, that is, the task-specific parameter is the parameter of each layer of the neural network. The function f(·) represents the shared intermediate conversion function, g i (·) denotes a task-specific transformation function.

[0157] Use S1 data to train the multi-step prediction model. During model training, use the gradient adjustment mechanism and dynamic weight allocation method to reduce the gradient conflict between multiple tasks, automatically adjust the relative importance of each task in the training process, optimize the performance of the overall multi-step prediction model, and finally obtain a trained multi-step prediction model. The specific implementation process is as follows:

[0158] (1) Figure 6 As shown in the figure, during the backward transfer of the gradient, the gradient is projected to reduce the gradient conflict between different tasks. Specifically, if the gradient directions of two tasks are opposite (i.e., conflict), these gradients need to be adjusted to make them more consistent. This adjustment is achieved by calculating the angle between the two gradient vectors and projecting one gradient onto the other. Figure 6 As an example, suppose there are two tasks with gradients of and The specific gradient adjustment process is described as follows:

[0159] First, calculate the angle between the two gradient vectors. If the angle is greater than 90 degrees (i.e. ), the two gradients are considered to be in conflict and gradient descent can be performed directly; if the angle is less than or equal to 90 degrees, it means that there is no gradient conflict and no adjustment is required. For two conflicting gradients, one of the gradients (for example ) to align with the other gradient No longer conflicting. Adjusted gradient The expression is as follows:

[0160]

[0161] Finally, the adjusted gradient is obtained and original Afterwards, these gradients will be used to update the parameters of the multi-step prediction model. The adjusted gradients help improve the performance of the multi-step prediction model in multi-task learning by reducing gradient conflicts between tasks, especially when multi-step predictions are involved. In the actual training and optimization process, the application of this method can effectively improve the coordination and overall performance of the multi-step prediction model between different prediction tasks.

[0162] The algorithm for updating the parameters of the multi-step prediction model using the adjusted gradient is implemented as follows:

[0163] The parameters of the multi-step prediction model that need to be updated include the weight matrix and the bias matrix. The set of tasks with different prediction steps is Set the loss function of the n-th step prediction to M n (θ), first loop to calculate the gradient of the n-th step prediction Then the gradients of different step lengths are traversed through two layers of loops, that is, the gradient predicted at a fixed nth step After that, for all the gradients on the prediction steps other than step n In chronological order and Compare, if the inner product of the two gradients is less than zero, then calculate exist projection on and from Subtract this projection from the gradient to adjust This process is repeated on the gradients of all tasks. Finally, these adjusted gradients are aggregated and used to update the parameters of the multi-step prediction model.

[0164] (2) Set the total number of training times to E. In order to more effectively coordinate the prediction tasks at different time steps, a dynamic weight allocation method is used, which can be expressed as the following formula:

[0165]

[0166]

[0167] In the formula, e represents the e-th round of training, M n (e-1) represents the prediction loss of the prediction of the nth step in the e-1th round, M n (e-2) represents the prediction loss of the prediction of the nth step in the e-2th round, w n (e-1) represents the ratio of the two. n (e) is the normalized weight of the n-th step prediction in all time-step tasks during the e-th round of training. τ is a hyperparameter that controls the sensitivity of weight distribution. The normalization operation of weights is to ensure that ∑ α λ α(e) = 1, which can ensure that the sum of the weights of each task in the multi-step prediction model training is always the same, thereby maintaining the stability of the multi-step prediction model training.

[0168] The loss function M(e) of the multi-step prediction model during the e-th round of training can be written as:

[0169]

[0170] Where L is the total number of samples in the data obtained by S1. n,l represents the true value of the lth sample in the nth step prediction task, Represents the predicted value of the lth sample in the nth step prediction task.

[0171] S6: Obtain real-time industrial process production data, pre-process it using the method of S1, and input it into the trained prediction model to obtain the prediction results.

[0172] like Figure 7 As shown, an embodiment of the present invention provides a coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding, and the device includes a memory and one or more processors. The memory stores executable code. When the one or more processors execute the executable code, they are used to implement the coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding in the above embodiment.

[0173] The coordinated optimization multi-step prediction device based on space-time coding-task decoding can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a computer or other device. The coordinated optimization multi-step prediction device based on space-time coding-task decoding can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located, reading the corresponding computer program instructions in the non-volatile memory into the internal memory for execution. From a hardware perspective, if Figure 7 As shown, it is a hardware structure diagram of any device with data processing capability where a coordinated optimization multi-step prediction device based on spatiotemporal coding-task decoding of the present invention is located, except Figure 7 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus of the present invention is located in the embodiments may also include other hardware, which will not be described in detail herein, based on the actual functions of the device with data processing capabilities.

[0174] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0175] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only exemplary, and the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0176] An embodiment of the present invention further provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding in the above embodiment is implemented.

[0177] The computer-readable storage medium may be an internal storage unit of any device with data processing capability in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device of any device with data processing capability, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capability. The computer-readable storage medium is used to store computer programs and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0178] The following is a specific industrial case of a sintering process to verify the effectiveness of the present invention. The data of this case was obtained from a sintering plant between December 8 and December 29, 2021, with a collection interval of one minute. Six input features were selected from the material variables and process variables in the sintering plant database, for a total of 12 feature input variables (ie, the number of channels C = 12). After preprocessing the original data, 31,049 data were obtained, the time step was set to 60 time steps (T = 60), and the prediction range was 7 steps (ie, the maximum prediction step N = 7). The data set is divided into a training set, a validation set, and a test set in chronological order, where the training set contains the first 24,839 data, the validation set contains the next 3,104 data, and the last 3,106 data are assigned to the test set. The training set and validation set are used to train the prediction model, and the test set is used as an online sample to detect the model training effect.

[0179] In order to illustrate the prediction effect of the method of the present invention, other existing prediction methods are also used for prediction. The models used for comparison include: AutoRegressive Integrated MovingAverage (ARIMA), Long Short Term Memory (LSTM), Long-and Short-term Time-series network (LSTNet), and Time Attention-Long Short Term Memory (TA-LSTM). Model evaluation indicators include root mean square error (RMSE), coefficient of determination (TSD), and ) and Mean Absolute Percentage Error (MAPE). The prediction effects of each model on online samples (test set data) are shown in Table 1:

[0180] Table 1 Comparison of multi-step prediction effects of various methods on real sintering datasets

[0181]

[0182] It can be seen from the results in Table 1 that the present invention shows the highest performance in all evaluation indicators, and the prediction accuracy of the 7th step and the average is significantly higher than that of other methods, thereby confirming the effectiveness and superiority of the present invention in predicting the multi-step sintering endpoint.

[0183] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the invention.

Claims

1. A coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding, characterized in that: The following steps are involved: S1: extracting historical industrial process production data within a period of time from the database of the sintering plant, performing data preprocessing on the data, and obtaining preprocessed historical industrial process production data; S2: Construct a spatiotemporal encoder composed of multiple spatial attention enhanced temporal convolution modules stacked together. Each spatial attention enhanced temporal convolution module includes a causal dilation convolution module and a spatial attention module. The input x t The output of the latter is connected through a one-dimensional convolution residual; the causal expansion convolution stacking module has several layers, and the input data of each layer passes through the causal expansion convolution layer to obtain its temporal characteristics and expand the receptive field, and is activated by the nonlinear activation function layer after weight normalization, and then outputted through the discard layer, and finally the output feature map O of time step t is obtained. t ;O t Input the spatial attention module, after the global attention layer extracts the global information of each channel, input it into the first fully connected layer, activated by the PReLU function, and the result is input into the second fully connected layer, activated by the sigmoid function, and outputs the output feature map z after spatial attention adjustment t ; S3: Perform fully adaptive noise set empirical mode decomposition on the historical sequence of target values ​​to obtain multiple intrinsic mode functions and a residual component; input the intrinsic mode functions with variance contribution ratio greater than the threshold into the gated neural network to extract time series features and generate a hidden state for each time point; input the hidden state into the fully connected layer and convert it into a dimension and format suitable for task-specific decoder processing; S4: Build a task-specific decoder and use the temporal attention mechanism to adaptively extract relevant information of the entire spatiotemporal encoder output in each prediction step to capture hidden states that are more closely related to the predicted target; In the final output layer, memory units with independent weights are used, and each prediction step has an independent set of parameters; the multi-step prediction results are finally output; S5: The spatiotemporal encoder and task-specific decoder form a multi-step prediction model, build a parameter sharing mechanism, and use the multi-task learning paradigm to train the multi-step prediction model; the training input is the data obtained in S1, and the gradient adjustment mechanism and dynamic weight allocation method are used during training to reduce the gradient conflict between multiple tasks and obtain a trained multi-step prediction model; S6: Obtain real-time industrial process production data, pre-process it using the method of S1, and input it into the trained multi-step prediction model to obtain the prediction results.

2. The coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding according to claim 1 is characterized in that: The S1 is specifically implemented through the following sub-steps: (1.1) Acquiring production data within a period of time from a time series database of a sintering plant as raw data, wherein the production data includes material variables and process variables; (1.2) Use the isolation forest method to detect outliers in the original data. For abnormal data, execute step (1.3); for the remaining data, jump to step (1.4); (1.3) Use local linear regression method to process abnormal data to maintain the continuity of time series data; (1.4) Integrate the normal data and the data obtained in step (1.3), and use wavelet decomposition denoising technology to filter out the noise; (1.5) Apply the sliding window method to the time series data obtained in step (1.4) to generate a set of time series data with a specific time step and prediction range, that is, the preprocessed historical production data.

3. The coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding according to claim 1 is characterized in that: In S2, the causal expansion convolution layer obtains the temporal feature F(t) in the input data through causal convolution, which is expressed as: In the formula, w represents the weight vector, x represents the input data, and w i represents the i-th component of w, x t-d*i represents the input data of the td*ith time step; k represents the convolution kernel size, t represents the current time step; d represents the expansion coefficient; The calculation expression of the weight normalization layer is as follows: Where W′ represents the standardized weight, g represents the learnable scaling parameter, and W d represents the original dilated convolution weight, ||W d || represents the L2 norm of the original dilated convolution weights; When the causal expansion convolution stack module has two layers, its input-output relationship expression is as follows: In the formula, represents the output feature map of the causal expansion stacking module at time step t, represents the input feature vector of the causal expansion stacking module at time step t, C is the number of channels, and T is the time step; * represents the causal expansion convolution operation; represents the normalized causal dilated convolution weight matrix of the first causal dilated convolution layer, represents the normalized causal dilated convolution weight matrix of the second causal dilated convolution layer; PReLU(·) represents the parameterized linear unit activation function; m1 and m2 are two randomly generated mask vectors used to implement the dropout operation; ⊙ represents the Hadamard product, which is used to apply the mask vector to the output of the causal dilated convolution layer.

4. The coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding according to claim 1 is characterized in that: In S2, the input of the spatial attention module is the output of the causal expansion convolution stack module O t , O t After the global information of each channel is extracted by the global attention layer, it is input into the first fully connected layer and activated by the PReLU function. The result is input into the second fully connected layer and activated by the sigmoid function. The output feature map after spatial attention adjustment is output. The expression is as follows: z t =O t ⊙σ(W2δ(W1I)); In the formula, represents the global information of each channel c, c = 1, 2, 3, ..., C, C is the number of channels; O t，c represents the output of the causal expansion convolution stack module at channel c; T is the time step, I represents a vector with C dimensions, and the value corresponding to each dimension is I c ; W1 represents the weight matrix of the first fully connected layer, W2 represents the weight matrix of the second fully connected layer; σ represents the sigmoid activation function.

5. The coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding according to claim 1 is characterized in that: In S3, the expression for performing fully adaptive noise set empirical mode decomposition on the target value history sequence is as follows: In the formula, y represents the historical sequence of target values, represents the pth intrinsic mode function, M is the total number of intrinsic mode functions obtained by the fully adaptive noise ensemble empirical mode decomposition, and R(t) represents the residual component; The expression of the variance contribution ratio is as follows: In the formula, represents the variance of the pth intrinsic mode function, and Var(y) represents the total variance of the target sequence y.

6. The coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding according to claim 1 is characterized in that: In S3, the final hidden state ψ output by the gated neural network Q The input is a fully connected layer, which is linearly transformed and a nonlinear activation function is added to generate a new feature representation; the output of the fully connected layer is used as the initial state of the task-specific decoder, which includes an initial hidden state and initial cell state The intrinsic mode function of the input gated neural network is embedded in each gated neural network unit. and In this approach, the intrinsic modality is transformed into a dimension and format suitable for processing by a task-specific decoder.

7. The coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding according to claim 1 is characterized in that: The S4 is specifically as follows: The context feature matrix S output by the spatiotemporal encoder is t , divided into T context vectors s from the time dimension t-T+1 ,s t-T+2 , ..., s t , these context vectors are input one-to-one into each time step of the task-specific decoder, where corresponds to the input to the task-specific decoder s t-T+k In time step; The temporal attention value of the hidden state of each task-specific decoder at time step t is expressed as follows: In the formula, represents the temporal attention value of the k-th context vector at time step t, is the overall weight matrix, tanh represents the hyperbolic tangent activation function; c t-1 The weight matrix, c t-1 represents the hidden state of the task-specific decoder at time step t-1; For t-T+k The weight matrix, s t-T+k represents the output of the spatiotemporal encoder at time step t-T+k; is the bias term; express The weight value after softmax normalization at the same time step t, N represents the maximum prediction step length; The input θ for each prediction step t It is calculated by weighing the state variables of the corresponding step and the output of all time steps of the spatiotemporal encoder part, and is combined with the predicted label y of the previous time step before being input into the memory unit. t-1 Splicing, the expression is as follows: θ t =concatenate(u t ,y t-1 ); The independent weight memory units in the task-specific decoder generate the final multi-step prediction results as follows: In the formula, represents the forget gate of the mth memory unit at time step t, The weight matrix representing the input data of the forget gate, The weight matrix representing the hidden state of the previous memory unit of the forget gate, Represents the bias term of the forget gate; represents the input gate of the mth memory unit at time step t, The weight matrix representing the input data of the input gate, The weight matrix representing the hidden state of the previous memory cell of the input gate, represents the bias term of the input gate; represents the candidate cell state of the mth memory unit at time step t, The weight matrix representing the input data of the candidate cell states, The weight matrix representing the previous hidden state of the candidate cell state, A bias term representing the candidate cell state; represents the cell state of the mth memory unit at time step t, represents the cell state of the mth memory unit at time step t-1; o m t represents the output gate of the mth memory unit at time step t, The weight matrix representing the input data of the output gate, The weight matrix representing the hidden state of the previous memory unit of the output gate, represents the bias term of the output gate; h t represents the hidden state of the last memory unit at time step t, c t represents the cell state of the last memory unit of the decoder at time step t; represents the prediction result of the last memory unit at time step t; FC represents the fully connected layer, and each memory unit outputs h t After the fully connected layer, the predicted value corresponding to the prediction step length is finally obtained 8. The coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding according to claim 1 is characterized in that: In S5, the input data passes through the shared parameter layer and the task layer in sequence, the multi-step prediction model learns common features through parameter sharing in the shared parameter layer, and predicts the result of each step through independent memory units in the task layer, while maintaining the connection between different prediction steps through the flow of time series state information; The gradient adjustment mechanism is as follows: Set the loss function of the n-th step prediction to M n (θ), loop to calculate the gradient of the prediction of the nth step The gradient of all other prediction steps In chronological order and Compare and determine whether the inner product of the two gradient vectors is less than 0. If not, there is no gradient conflict between the two gradient vectors and no adjustment is required; if so, there is a gradient conflict between the two gradient vectors and the calculation is exist projection on and from This projection is subtracted from the gradient to adjust The dynamic weight allocation method is specifically as follows: set the number of training times to E, calculate the normalized weight of the n-th step prediction in all time step tasks in each round of training, n = 1, 2, 3, ..., N, N is the time step, and the expression is as follows: Where M n (e-1) represents the prediction loss of the prediction of the nth step in the e-1th round, e = 1, 2, 3, ..., E; M n (e-2) represents the prediction loss of the prediction of the nth step in the e-2th round, w n (e-1) represents the ratio of the two; λ n (e) is the normalized weight of the n-th step prediction in all time-step tasks during the e-th round of training; τ is a hyperparameter; The loss function M(e) of the multi-step prediction model in the e-th round of training is: Where L is the total number of samples in the preprocessed historical industrial process production data; y n,l represents the true value of the lth sample in the nth step prediction task, Represents the predicted value of the lth sample in the nth step prediction task.

9. A coordinated optimization multi-step prediction device based on target-related spatiotemporal coding-task decoding, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, they are used to implement the coordinated optimized multi-step prediction method based on target-related spatiotemporal coding-task decoding as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the coordinated optimization multi-step prediction method based on target-related spatiotemporal coding-task decoding as described in any one of claims 1 to 8 is implemented.

Images