Multi-scale time sequence prediction method and device and application of multi-scale time sequence prediction method and device in dust removal fan control

Abstract

The invention discloses a multi-scale time sequence prediction method and device and application of the multi-scale time sequence prediction method and device in dust removal fan control, and relates to the technical field of artificial intelligence. According to the scheme, features of different granularities are extracted through multi-scale decomposition, bidirectional information interaction between coarse and fine granularities is achieved in combination with an asymmetric cross-scale attention mechanism, and the method and device are high in practicability. Meanwhile, cross-scale time sequence feature fusion is realized through an asymmetric cross attention mechanism, noise can be effectively reduced, feature complementarity can be enhanced, through the technical scheme, short-term fluctuation and long-term trend in a time sequence can be captured at the same time, and the problem of feature missing caused by single scale in a traditional method is solved. The cross-scale attention mechanism promotes information fusion among the subsequences with different granularities, and the modeling capability of the model for the complex time dependency relationship is enhanced, so that the accuracy of a prediction result in a variable scene is improved, and the prediction precision and the industrial control effect are improved.

Description

[0001] This application claims the priority of the domestic application 202510873242.X, the filing date of which is June 27, 2025. TECHNICAL FIELD

[0002] The present application relates to the field of artificial intelligence, in particular to a multi-scale time series prediction method and device and its application in dust removal fan control. BACKGROUND

[0003] Time series prediction is a technology based on historical data values of time series to predict future values, which plays an important role in economic and financial markets, supply chain management, transportation, energy consumption, weather and industrial production, etc. In these fields, time series prediction is usually used as an auxiliary tool for decision-making, which can effectively reduce costs and improve efficiency.

[0004] Traditional time series models can only analyze and predict at a single particle scale, and do not fully exploit the complementarity of multi-scale time series data, making it difficult to effectively capture long-term dependencies in time series information at different scales, resulting in insufficient prediction accuracy. When time series prediction technology is applied to industrial process control, such as dust removal systems in steel plants and chemical plants, the above problem of insufficient prediction accuracy will further cause response delay of the dust removal system, leading to reduced dust removal efficiency and endangering the health of on-site personnel due to the high temperature, high pressure and super-long ventilation duct conditions of the running environment. SUMMARY

[0005] The present application provides a multi-scale time series prediction method, device and its application in dust removal fan control, which is used to achieve the purpose of improving time series prediction accuracy and further optimizing the effect of industrial process control.

[0006] To achieve the above-mentioned purpose of the application, the first aspect of the present application provides a multi-scale time series prediction method, comprising:

[0007] obtaining an original time series;

[0008] performing sequence decomposition on the original time series to obtain a plurality of sub-time series of different particle scales;

[0009] based on the order of the particle scale thickness of each sub-time series, the first two sub-time series are sequentially taken as a first sequence and a second sequence, and then based on the first sequence and the second sequence, QKV features and attention weights corresponding to the QKV features are calculated, wherein the Q feature is calculated based on the first sequence, and the K feature and the V feature are calculated based on the second sequence feature;

[0010] perform feature fusion processing on the second sequence and the attention weight to obtain a fused time sequence, if a next sub-time sequence corresponding to the current second sequence is not empty, update the first sequence based on the fused time sequence, update the second sequence based on the next sub-time sequence, and obtain a latest fused time sequence based on the updated first sequence and second sequence, if the next sub-time sequence corresponding to the current second sequence is empty, construct a multi-scale time sequence according to the first sub-time sequence and each fused time sequence;

[0011] perform time series prediction based on the multi-scale time sequence to obtain a prediction result.

[0012] Preferably, it further comprises:

[0013] perform nonlinear transformation preprocessing on the original time sequence by a preset multi-layer perception.

[0014] Preferably, the multi-layer perception comprises a layer normalization module, a fully connected layer, a GELU activation function, a Dropout calculation module, and a residual connection processing module.

[0015] Preferably, the sequence decomposition of the original time sequence to obtain a plurality of sub-time sequences of different particle sizes comprises:

[0016] perform sequence decomposition on the original time sequence by an average pooling downsampling processing method to obtain a plurality of sub-time sequences of different particle sizes.

[0017] Preferably, the feature fusion processing based on the second sequence and the attention weight to obtain a fused time sequence comprises:

[0018] perform layer normalization processing on the sum of the second sequence and the attention weight to obtain a preliminary fusion result;

[0019] perform high-order feature extraction on the preliminary fusion result by a preset feedforward neural network, and then sequentially perform residual connection and layer normalization processing on the extracted high-order feature and the preliminary fusion result to obtain a fused time sequence.

[0020] Preferably, the time series prediction based on the multi-scale time sequence to obtain a prediction result comprises:

[0021] input the multi-scale time sequence into a preset linear neural network to determine a prediction result based on an output result of the linear neural network.

[0022] Preferably, the determination of a prediction result based on the output result of the linear neural network comprises:

[0023] According to the multi-scale prediction result sequence output by the linear neural network, each element in the multi-scale prediction result sequence is weighted and summed in combination with a preset learnable weight coefficient sequence, so as to obtain a final prediction result according to a sum result.

[0024] The second aspect of the application provides a multi-scale time series prediction device, comprising:

[0025] An original sequence acquisition unit is configured to acquire an original time series.

[0026] A multi-scale sequence decomposition unit is configured to perform sequence decomposition on the original time series to obtain a plurality of sub-time series of different particle scales.

[0027] A cross-scale attention processing unit is configured to sequentially take the first two sub-time series as a first sequence and a second sequence based on the order of particle scale thickness of each sub-time series, and then calculate QKV features and attention weights corresponding to the QKV features based on the first sequence and the second sequence, wherein the Q features are calculated based on the first sequence, and the K features and the V features are calculated based on the second sequence features.

[0028] A cross-scale feature fusion unit is configured to perform feature fusion processing on the second sequence and the attention weights to obtain a fusion time series, and if the next sub-time series corresponding to the current second sequence is not empty, update the first sequence based on the fusion time series, and update the second sequence based on the next sub-time series, so as to obtain the latest fusion time series based on the updated first sequence and second sequence, and if the next sub-time series corresponding to the current second sequence is empty, construct a multi-scale time series according to the first sub-time series and each fusion time series.

[0029] A time series prediction unit is configured to perform time series prediction based on the multi-scale time series to obtain a prediction result.

[0030] The third aspect of the application provides a dust removal fan operation control method, comprising:

[0031] Collecting working condition data of the dust removal fan at a plurality of time points;

[0032] Taking the working condition data at the plurality of time points as an original time series, and obtaining a working condition data prediction result of the dust removal fan at a preset future time according to the multi-scale time series prediction method provided in the first aspect of the application;

[0033] According to the working condition data prediction result, in combination with the correlation between the working condition data and the operation frequency of the dust removal fan, determining a target operation frequency corresponding to the working condition data prediction result;

[0034] The target operating frequency is transmitted to the frequency converter so that the frequency converter adjusts the fan speed according to the target operating frequency.

[0035] The fourth aspect of this application provides a dust removal system, including: a dust removal fan, a frequency converter, a control module, and several sensing components;

[0036] The sensing component is installed in the dust removal fan and is used to collect the operating condition data of the dust removal fan at multiple time points.

[0037] The control module is communicatively connected to the frequency converter and each of the sensing components, and the signal output terminal of the frequency converter is communicatively connected to the control terminal of the dust removal fan.

[0038] The control module includes a memory and a processor. The memory is used to store program code, which corresponds to the dust collector fan operation control method provided in the third aspect of this application. The processor is used to read and execute the program code to implement the dust collector fan operation control method.

[0039] As can be seen from the above technical solutions, this application has the following advantages:

[0040] The solution provided in this application extracts features of different granularities through multi-scale decomposition, and combines an asymmetric cross-scale attention mechanism to achieve bidirectional information interaction between coarse and fine granularities. Simultaneously, it achieves cross-scale temporal feature fusion through an asymmetric cross-attention mechanism, effectively reducing noise and enhancing feature complementarity. Through these technical solutions, this application can simultaneously capture short-term fluctuations and long-term trends in time series, solving the feature loss problem caused by the single scale in traditional methods. The cross-scale attention mechanism promotes information fusion between subsequences of different granularities, enhances the model's ability to model complex temporal dependencies, and thus improves the accuracy of prediction results in variable scenarios. This effectively solves the problem that traditional methods cannot capture long-term dependencies, thereby improving prediction accuracy and enabling more precise predictions to improve control performance in industrial scenarios. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating an embodiment of a multi-scale time series prediction method provided in this application.

[0043] Figure 2 This is a flowchart illustrating the processing logic of a multilayer perceptron in a multi-scale temporal prediction method provided in this application.

[0044] Figure 3 The flowchart illustrates the fusion logic of multi-scale subsequences in a multi-scale time series prediction method provided in this application.

[0045] Figure 4 This is a schematic diagram of an embodiment of a multi-scale time series prediction device provided in this application.

[0046] Figure 5 This is a flowchart illustrating an embodiment of a dust removal fan operation control method provided in this application.

[0047] Figure 6 This is a schematic diagram of the structure of a control module embodiment in a dust removal system provided in this application.

[0048] Figure 7 This is a comparison chart showing the prediction results of pressure value working condition data obtained based on the multi-scale time series prediction method provided in this application.

[0049] Figure 8 This is a comparison chart showing the prediction results of dust volume operating condition data obtained based on the multi-scale time series prediction method provided in this application.

[0050] Figure 9 This is a comparison chart showing the pre-prediction performance of temperature condition data obtained based on the multi-scale time-series prediction method provided in this application.

[0051] Figure 10 This is a comparison chart showing the prediction performance of stress value working condition data obtained based on existing deep learning prediction methods.

[0052] Figure 11 This is a comparison chart showing the prediction performance of dust volume data based on existing deep learning prediction methods.

[0053] Figure 12 This is a comparison chart showing the prediction performance of temperature condition data obtained based on existing deep learning prediction methods. Detailed Implementation

[0054] In existing technologies, time series forecasting is widely used in fields such as economics and finance, supply chain management, and energy consumption. Traditional methods are usually based on a single time scale, making it difficult to effectively capture dynamic characteristics at different time granularities. For example, in the scenario of electricity load forecasting, short-term fluctuations and long-term trends are often intertwined, and single-scale models cannot simultaneously take into account hourly changes and quarterly periodic patterns, leading to forecast results that deviate from actual needs.

[0055] To address the aforementioned issues, the inventors discovered that the insufficient ability of traditional models to fuse multi-scale temporal features is a key factor limiting prediction accuracy. By analyzing the hierarchical characteristics of temporal data in the target scene, they realized that cross-scale interactions between subsequences of different scales may contain complementary information. Further research revealed that fusing the global trend of coarse-grained sequences with the local details of fine-grained sequences using an attention mechanism can enhance the model's ability to capture long-term dependencies.

[0056] In view of this, embodiments of this application provide a multi-scale time series prediction method, apparatus and its application in dust removal fan control, to solve the technical problem of insufficient accuracy in existing time series prediction.

[0057] To make the inventive objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0058] First, a detailed description of an embodiment of a multi-scale time series prediction method provided in this application is as follows:

[0059] Please see Figure 1 This application provides a multi-scale time series prediction method, comprising:

[0060] Step 101: Obtain the original time series;

[0061] Step 102: Decompose the original time series to obtain multiple sub-time series with different particle scales;

[0062] Step 103: Based on the order of the granularity of each sub-time series, take the first two sub-time series as the first and second series respectively, and then calculate the QKV features and the attention weights corresponding to the QKV features based on the first and second series.

[0063] Among them, the Q feature is calculated based on the first sequence, while the K and V features are calculated based on the second sequence features;

[0064] Step 104: Perform feature fusion processing based on the second sequence and attention weights to obtain a fused time series. If the next sub-time series corresponding to the current second sequence is not empty, then update the first sequence based on the fused time series and update the second sequence based on the next sub-time series. Based on the updated first and second sequences, obtain the latest fused time series. If the next sub-time series corresponding to the current second sequence is empty, then construct a multi-scale time series based on the first sub-time series and each fused time series.

[0065] Step 105: Perform time series prediction based on multi-scale time series to obtain prediction results.

[0066] The original time series refers to the set of observation data arranged in chronological order, which can be collected through sensors, databases, or log systems to provide basic prediction input. Sequence decomposition refers to converting the original sequence into multiple subsequences with different temporal granularities. This can be achieved using methods such as average pooling downsampling, adaptive multi-scale decomposition, max or min pooling, and multi-resolution convolution. Its purpose is to separate feature patterns at different time scales. Granularity scale order refers to arranging the subsequences according to the size of the time window, which can be achieved by comparing the downsampling factors of the pooling layers, used to establish a processing order from coarse-grained to fine-grained or vice versa.

[0067] In QKV feature computation, Q features are generated based on the first sequence, while K and V features are generated based on the second sequence. This can be achieved through a linear transformation matrix, which serves to establish query-key pairs with cross-scale attention interaction. Feature fusion processing involves weighted superposition of attention weights with the second sequence, which can be achieved through residual connections and layer normalization. This preserves the original features and enhances cross-scale information transfer. For example, the multi-scale time series after decomposition is... First, the two time series at different scales are... , As input to the asymmetric cross-scale attention module, the sequence After linear projection, it is mapped to a space, generating a query matrix. (Query); sequence After linear projection, the matrix is ​​mapped to two different spaces, generating bond matrices respectively. (Key) and value matrix (Value). Then, the obtained... The feature matrix is ​​used to calculate the corresponding attention weights.

[0068] Specifically, the original time series is downsampled and decomposed into high-frequency, mid-frequency, and low-frequency subsequences. Based on the granularity of each subsequence, the coarsest or finest granular subsequence is designated as the first sequence, and the next coarsest as the second sequence. Using an asymmetric attention mechanism, the first sequence generates a Q-vector to query the K and V vectors of the second sequence, calculating cross-scale correlation weights. This weight is then superimposed on the second sequence. If any unfused subsequences remain, the current fusion result is used as the new first sequence, and the next subsequence is used as the new second sequence for further iteration. Finally, the coarsest or finest granular subsequence is combined with all fusion results to form a multi-scale time series containing multi-level features, which is then input into the prediction module as the output.

[0069] This approach extracts features of different granularities through multi-scale decomposition and combines an asymmetric cross-scale attention mechanism to achieve bidirectional information interaction between coarse and fine granularities. Simultaneously, it achieves cross-scale temporal feature fusion through an asymmetric cross-attention mechanism, effectively reducing noise and enhancing feature complementarity. Through these technical solutions, this application can simultaneously capture short-term fluctuations and long-term trends in time series, solving the feature loss problem caused by the single scale in traditional methods. The cross-scale attention mechanism promotes information fusion between subsequences of different granularities, enhancing the model's ability to model complex temporal dependencies, thereby improving the accuracy of prediction results in variable scenarios.

[0070] Based on the above basic embodiments, the method provided in this application further proposes to perform nonlinear transformation preprocessing on the original time series by a preset multilayer perceptron before executing step 102.

[0071] More specifically, the multilayer perceptron includes a layer normalization module, a fully connected layer, a GELU activation function, a Dropout calculation module, and a residual connection processing module. The overall formula is as follows:

[0072]

[0073] In the formula, It is the output of the multilayer perceptron. It is the input of the multilayer perceptron (the original time series). It is the output after layer normalization. It is a weight matrix. It is a bias term.

[0074] The layer normalization module is a unit that standardizes the input time series features. Specifically, it can be implemented by calculating the mean and variance along the feature dimension to eliminate the dimensional differences of the input features and enhance the stability of the training process. The formula is as follows:

[0075]

[0076] In the formula, It is the mean. It is variance. It is usually a very small constant used to prevent the denominator from being zero. It is a learnable scaling factor. It is the translation parameter.

[0077] A fully connected layer is a neural network layer that performs a linear transformation. Specifically, it can be implemented by multiplying the weight matrix by the input vector and adding bias terms. This maps standardized temporal features to a new dimensional space to extract latent information. The formula is shown below, where... It is the output of the fully connected layer.

[0078]

[0079] The GELU activation function is an activation function based on Gaussian error linear units. Its nonlinear transformation characteristics can be achieved through approximate calculation methods to enhance the network's ability to fit complex temporal patterns.

[0080] The Dropout module is a unit that randomly masks the outputs of some neurons during the training phase. This can be achieved by using a Bernoulli distribution to generate the mask matrix, reducing the co-adaptability between neurons and suppressing overfitting. Dropout randomly "dormants" some neurons with a specific probability, meaning they do not participate in the current forward and backward propagation processes. During the testing phase, to obtain the most accurate prediction results, Dropout reactivates all neurons and uses the average value of the fully connected layers to make a comprehensive judgment, ensuring the model can handle new, unseen data in its optimal state, achieving stable and reliable performance output. The residual connection processing module is a computational unit that directly adds the module's input and output, specifically using element-wise tensor addition. This alleviates the gradient vanishing problem in deep network training, not only accelerating the training speed but also significantly improving training accuracy. This allows deep networks to successfully learn richer and more complex feature representations, providing a solid architectural foundation for solving highly complex tasks.

[0081] Specifically, such as Figure 2As shown, after the original time series is input into a multilayer perceptron, it first undergoes layer normalization to eliminate differences in feature dimensions. Then, a linear transformation is performed through fully connected layers to achieve feature interaction. After introducing nonlinear computation using the GELU activation function, Dropout (random deactivation) is used to randomly disable some neurons to enhance generalization performance. Finally, residual connections are used to superimpose the processed features with the original input, forming temporal features with high-order expressive power. This processing method allows for explicit modeling of potential correlation patterns between time steps, while mapping the original feature space to a representation space more suitable for multi-scale decomposition.

[0082] This scheme, by introducing a nonlinear transformation module, effectively extracts complex patterns such as periodicity and trends implicit in time series data. It achieves effective mixing and enhancement of time series features in the transformation space, providing more discriminative input features for subsequent multi-scale decomposition. Simultaneously, through combined layer normalization, Dropout, and residual connections, it effectively controls the stability and generalization ability of the training process while preserving the feature interaction capabilities of fully connected layers. This preprocessing mechanism can capture dynamic patterns in the original data that are difficult to extract using linear methods, achieving a robust mapping of time series features in the nonlinear space, and providing a high-information-density feature representation foundation for subsequent cross-scale attention fusion.

[0083] Furthermore, regarding the sequence decomposition process in step 102 of this application, the specific steps are as follows: when performing sequence decomposition on the original time series, an average pooling downsampling method is used to obtain multiple sub-time series with different granular scales.

[0084] The average pooling downsampling method refers to dividing the original time series into segments with a fixed window length and taking the average value of the data within each window as the new data point. Specifically, it can be implemented by using a moving average operation with a window length of 2, generating subsequences with different time resolutions by progressively reducing the sampling rate.

[0085] The granular scale refers to the degree of difference in temporal resolution between sub-time series, which can be controlled by adjusting the step size parameter of the pooling window. For example, setting an exponential growth mode with a step size of 2 will make the time step of each subsequent sub-series twice that of the previous one.

[0086] Specifically, the process begins by inputting the original time series data into the downsampling processing module, where multi-level average pooling operations are used to generate subsequence sets with different time resolutions. For example, for an original sequence of length T, the first-level pooling window is set to 2, resulting in a subsequence of length T / 2; the second-level pooling window is set to 4, resulting in a subsequence of length T / 4. This hierarchical processing preserves high-frequency detail information in the fine-grained subsequences, while the coarse-grained subsequences capture long-term trend features.

[0087] This scheme directly reduces the dimensionality of time-domain data through average pooling downsampling, which not only improves computational efficiency but also effectively suppresses noise interference through the local averaging characteristics of the sliding window. This is more conducive to capturing long-term dynamic patterns in time series. Through the above technical solution, the original time series can be decomposed into multi-scale subsequences with complementary information. The fine-grained subsequences retain short-term fluctuation details, while the coarse-grained subsequences represent overall trend changes. This effective separation of multi-scale features provides a foundation for subsequent cross-scale attention mechanisms.

[0088] Furthermore, regarding the feature fusion processing based on the second sequence and attention weights proposed in step 105 of this application to obtain the fused time series, the process can be seen in the following example: The sum of the second sequence and attention weights is subjected to layer normalization to obtain a preliminary fusion result; high-order features are extracted from the preliminary fusion result using a pre-defined feedforward neural network; then, the extracted high-order features and the preliminary fusion result are sequentially subjected to residual connections and layer normalization to obtain the fused time series. For a time series decomposed into n scales, n-1 asymmetric cross-scale attention modules need to be constructed sequentially, each module processing the output of the previous module and the next time scale. By continuously repeating the above cross-scale attention mechanism and feedforward neural network process, it is ensured that the last scale can be processed. Finally, the multi-scale time series after the asymmetric cross-scale attention modules is used for subsequent multi-scale prediction.

[0089] Layer normalization refers to the standardization of feature data, which can be achieved by calculating the mean and variance and performing a normalization transformation. Its purpose is to eliminate differences in feature dimensions and improve training stability. Feedforward neural networks are multi-layered structures composed of linear layers and activation functions. Specifically, they can be implemented using a combination of fully connected layers and the GELU activation function. Their purpose is to extract high-order temporal features through nonlinear transformations. Residual connections are operations that add the network input and output. Specifically, they can be implemented using skip connection structures. Their purpose is to alleviate the vanishing gradient problem and accelerate model convergence. Asymmetric cross-scale attention modules are attention mechanisms that allow only coarse-grained sequences to generate query vectors while fine-grained sequences generate key-value vectors. Specifically, they can be implemented by using sub-sequences of different scales to generate Q, K, and V features respectively. Their purpose is to establish cross-scale temporal dependencies.

[0090] like Figure 3 As shown, assume that the multi-scale time series after downsampling decomposition is ,in Represents the sequence at the original scale. This represents the coarsest-grained scale of the sequence. First, two time series at different scales are... , As input to the asymmetric cross-scale attention module, the sequence After linear projection, it is mapped to a space, generating a query matrix. (Query); sequence After linear projection, the matrix is ​​mapped to two different spaces, generating bond matrices respectively. (Key) and value matrix (Value). Then, the obtained... The matrix is ​​used to calculate the attention weights. Then, the attention is output. With input scale Perform residual connections and obtain the fused representation through layer normalization. Finally, the fusion results will be... High-order features are extracted using feed-forward networks (FFNs), and residual connections and layer normalization operations are performed to obtain the final output time series representation. The calculation process is as follows:

[0091]

[0092]

[0093]

[0094]

[0095] in, It is a learnable weight matrix. It is the scaling factor.

[0096] For a time series decomposed into n scales, it is necessary to construct n-1 asymmetric cross-scale attention modules, each module being used to process the output of the previous module. With the next time scale By repeatedly employing the aforementioned cross-scale attention mechanism and feedforward neural network process, the final scale is ensured. This allows for processing. Finally, these multi-scale time series, after passing through the asymmetric cross-scale attention module, are represented as follows: It is used for subsequent multi-scale prediction.

[0097] Specifically, in a single feature fusion process, the second sequence is first added to the attention weights and then input into a normalization module to eliminate feature distribution bias. Subsequently, a feedforward neural network performs a non-linear transformation on the normalization result, for example, using a structure with a GELU activation function inserted between two fully connected layers to extract a time series representation containing high-order dynamic features. The feedforward network output is then residually connected to the initial fusion result, and a second layer normalization is performed to generate the final fused time series. For decomposition results involving multiple scales, multiple asymmetric cross-scale attention modules need to be constructed sequentially. For example, when the original sequence is decomposed into three scales, two modules need to be constructed to handle the cross-scale interactions of X1 and X2, and O1 and X3 respectively, and so on. Each module uses the same processing flow, gradually fusing the fused temporal features of all scales through a chain structure.

[0098] This scheme ensures feature distribution stability by introducing layer normalization, enhances feature representation by utilizing feedforward neural networks, and improves gradient propagation efficiency by combining residual connections. The asymmetric attention mechanism effectively distinguishes the roles of sequences at different scales in feature extraction; for example, coarse-grained sequences are better suited for capturing global patterns as query benchmarks, while fine-grained sequences provide local details as key-value criteria. This structured feature fusion approach can more accurately establish multi-scale temporal correlations compared to traditional methods. Through the above technical solutions, this application solves the problem of information loss caused by insufficient cross-scale feature fusion in existing technologies and avoids feature distribution conflicts caused by simple addition or concatenation operations. The combined use of layer normalization and residual connections effectively improves the stability of model training and prevents training oscillations caused by gradient anomalies. The asymmetric attention mechanism strengthens the targeting of cross-scale interactions through role differentiation, enabling coarse-grained trend features and fine-grained fluctuation features to participate in prediction in a complementary manner. The introduction of feedforward neural networks can further explore nonlinear temporal patterns in the fused features, providing highly discriminative multi-scale feature representations for subsequent predictions.

[0099] Furthermore, step 106 of this application proposes a time series prediction based on multi-scale time series, and the prediction results include: inputting the multi-scale time series into a preset linear neural network, and determining the prediction results based on the output of the linear neural network; and performing a weighted summation of each element in the multi-scale prediction result sequence based on the multi-scale prediction result sequence output by the linear neural network and a preset learnable weight coefficient sequence, so as to obtain the final prediction result based on the summation result.

[0100] Linear neural networks refer to mapping models composed of fully connected layers. Specifically, they can be implemented using a fully connected layer structure that matches the input dimension with the feature dimensions of multi-scale time series. Their function is to map multi-scale time series data. Each time step feature is mapped to a predicted value. A learnable weight coefficient sequence refers to a set of weight parameters that are dynamically adjusted through the backpropagation algorithm. Specifically, it can be implemented in the form of trainable vectors or matrices, and its function is to dynamically allocate the weights based on the contribution of prediction results at different time scales to the final result.

[0101] Specifically, after the multi-scale time series is input into a linear neural network, the features at each time scale are passed through fully connected layers to generate prediction result sequences for the corresponding scale. These predicted sequences are concatenated to form a multidimensional tensor, from which the weight coefficient sequence can be learned. The prediction results at each scale are weighted using matrix multiplication. During the weighting process, the prediction result at each time step is scaled according to the weight coefficient of its respective scale. Finally, the results are summed along the feature dimension to generate a single-dimensional output sequence as the final prediction result. The formula is as follows:

[0102]

[0103] Among them, the weights of each scale This can be automatically learned through model training, ensuring a balance between global and local trends in the fusion result. This module can effectively integrate prediction information from different time scales, enabling the final output to have both local mutation response capabilities and global trend awareness capabilities, significantly improving overall prediction accuracy and making it suitable for a variety of complex scenarios.

[0104] This solution achieves dynamic adaptive fusion of multi-scale prediction results by combining learnable weight coefficients with a linear neural network. It automatically adjusts the contribution ratio of each time scale to the final result based on the characteristics of the input data. Through this technical solution, this application solves the problem of rigid feature fusion methods at different scales in multi-scale time series prediction, effectively improving the model's ability to capture multi-scale correlation features in complex time series patterns. The dynamic weight allocation mechanism enables the model to autonomously identify key time scales and suppress noise interference, thereby improving the accuracy and robustness of the final prediction results.

[0105] To further demonstrate the technical effectiveness of the multi-scale time series prediction method provided in this application, simulation test results based on multiple time series data are presented in Table 1.

[0106] Table 1. Comparison of test results between the prediction method provided in this application and other existing methods.

[0107]

[0108] Eight publicly available datasets were used in the experiment. The meteorological dataset contained meteorological indicators, the solar energy dataset involved solar power generation data, the electricity and traffic datasets covered electricity consumption and traffic flow information, respectively, and the ETTh1, ETTh2, ETTm1, and ETTm2 datasets represented the operating status of power transformers. Smaller values ​​for MSE (Mean Squared Error) and MAE (Mean Absolute Error) indicate better model performance.

[0109] By comparing the performance of different time series prediction methods on different datasets, it can be seen that the method proposed in this application is superior to other existing methods.

[0110] The above is a detailed description of an embodiment of a multi-scale time series prediction method provided in this application. The following is a detailed description of an embodiment of a multi-scale time series prediction device corresponding to the above method embodiment.

[0111] Please see Figure 4 This application provides an embodiment of a multi-scale time series prediction device, comprising:

[0112] The original sequence acquisition unit 201 is used to acquire the original time series;

[0113] The multi-scale sequence decomposition unit 202 is used to decompose the original time series into multiple sub-time series with different particle scales.

[0114] The cross-scale attention processing unit 203 is used to take the first two sub-time series as the first sequence and the second sequence in turn based on the order of the coarseness of the granular scale of each sub-time series, and then calculate the QKV features and the attention weights corresponding to the QKV features based on the first sequence. The Q features are calculated based on the first sequence, while the K features and V features are calculated based on the features of the second sequence.

[0115] The cross-scale feature fusion unit 204 is used to perform feature fusion processing based on the second sequence and attention weights to obtain a fused time series. If the next sub-time series corresponding to the current second sequence is not empty, the first sequence is updated based on the fused time series, and the second sequence is updated based on the next sub-time series. The latest fused time series is obtained based on the updated first and second sequences. If the next sub-time series corresponding to the current second sequence is empty, a multi-scale time series is formed based on the first sub-time series and each fused time series.

[0116] The time series prediction unit 205 is used to perform time series prediction based on multi-scale time series and obtain prediction results.

[0117] The above is a detailed description of an embodiment of a multi-scale time series prediction device provided in this application. The following is a detailed description of an application embodiment of a multi-scale time series prediction method corresponding to the above method embodiment in a specific industrial dust removal fan operation control scenario.

[0118] Please see Figure 5 This embodiment proposes a dust removal fan operation control method, including:

[0119] Step 301: Collect operating data of the dust collector fan at multiple time points;

[0120] Step 302: Using operating data at multiple time points as the original time series, the operating data prediction results of the dust removal fan at a preset future time are obtained through a multi-scale time series prediction method.

[0121] Step 303: Based on the predicted results of the operating data, and combined with the correlation between the operating data and the operating frequency of the dust collector fan, determine the target operating frequency corresponding to the predicted results of the operating data; transmit the target operating frequency to the frequency converter so that the frequency converter can adjust the fan speed according to the target operating frequency.

[0122] Among them, operating condition data refers to key parameters reflecting the equipment status during the operation of the dust collector fan. Specifically, this data can be collected from sensors such as temperature, pressure, and dust volume to characterize the fan's real-time operating status. The multi-scale time series prediction method refers to the method mentioned in the above embodiments. This method uses a prediction model that performs multi-scale decomposition and cross-scale feature fusion on the original time series. Specifically, it can employ an algorithm structure including sequence decomposition, attention mechanisms, and feature weighted fusion to capture dynamic correlations at different time granularities. The target operating frequency refers to the motor drive parameters dynamically adjusted based on the prediction results. This can be calculated using a preset operating condition-frequency mapping table or regression model to optimize fan operating efficiency. The frequency converter is an electronic device used to adjust the motor speed. Specifically, it can employ an industrial controller with frequency modulation capabilities to achieve precise control of the fan speed.

[0123] Specifically, during the operation of the dust collector fan, sensors periodically collect operating condition data to form a raw time series. This series is input into a multi-scale time series prediction model for decomposition and cross-scale feature fusion to generate predicted operating conditions for future time points. The prediction results are matched with preset operating condition-frequency association rules to calculate the optimal operating frequency parameters. These frequency parameters are transmitted to the frequency converter via an industrial communication protocol, triggering the frequency modulation function of the motor drive circuit, thereby achieving closed-loop control of the fan speed. When the prediction model accurately captures the trend of operating condition changes, the system can adjust the operating frequency in advance, avoiding the response lag caused by prediction delays in traditional control methods.

[0124] This solution proactively identifies potential operating condition fluctuations through multi-scale time-series prediction and achieves active speed regulation via variable frequency control, addressing the issue of reduced dust removal efficiency caused by response delays in traditional methods under high temperature, high pressure, and ultra-long ventilation duct environments. Existing single-time-scale prediction models struggle to capture multi-dimensional correlation features under complex operating conditions. This solution effectively integrates time-series features of different granularities through a cross-scale attention mechanism, enhancing the predictive accuracy's support for control decisions. Through this technical solution, this application resolves the control delay problem caused by insufficient predictive accuracy in industrial dust removal scenarios, achieving dynamic matching of fan operating parameters with changes in operating conditions. This improves the response speed and operational stability of the dust removal system in complex industrial environments, while simultaneously reducing energy waste and equipment wear risks caused by control lag.

[0125] In some specific implementations, the data acquisition interval can be set to the second or minute level, for example, collecting duct pressure data every 30 seconds. The sequence decomposition in multi-scale time series prediction methods can employ sliding window average pooling, for example, decomposing the original sequence into three sub-sequences at the hour, ten-minute, and minute levels. The frequency adjustment response time of the frequency converter can be controlled at the millisecond level, for example, by using an intelligent frequency converter with an RS-485 communication interface.

[0126] It should be noted that in traditional steel plant dust removal systems, an extremely long ventilation duct exists between the dust collector fan and the blast furnace, resulting in a time lag between sensor readings and the fan's response. For example, the fan may only be activated or the airflow increased when the sensor detects a dust concentration exceeding a set threshold, by which time some dust has already escaped, indicating a significant time delay. Simultaneously, traditional dust collector fans often operate at a fixed frequency, switching between "fully on" and "fully off," a process that often leads to excessive energy waste. To address this issue, this embodiment employs the improved time-series prediction method mentioned in the previous embodiments. This proactive control strategy, by capturing fluctuations at different scales, can identify nonlinear and long-term dependencies that traditional control methods cannot respond to, predicting sensor readings over a future period. By comprehensively analyzing the predicted values, the operating frequency of the dust collector fan is adjusted in advance, resulting in higher dust removal efficiency and reduced energy consumption of the dust removal system.

[0127] Furthermore, this application proposes a dust removal system, including a dust removal fan, a frequency converter, a control module, and several sensing components. The sensing components are installed in the dust removal fan to collect operating data of the dust removal fan at multiple time points. The control module is communicatively connected to the frequency converter and each sensing component, and the signal output terminal of the frequency converter is communicatively connected to the control terminal of the dust removal fan. The control module includes a memory 33 and a processor 31, which are connected via a communication bus 34. The memory stores program code corresponding to the dust removal fan operation control method, and the processor reads and executes the program code to implement the dust removal fan operation control method. The specific architecture of the control module is as follows: Figure 6 As shown.

[0128] The sensing components refer to a set of sensors used to monitor the real-time operating status of the dust collector fan. Specifically, these can be implemented using temperature sensors, pressure sensors, or flow sensors, collecting operating data at different time points to form a raw time series. The communication connection between the control module and the frequency converter refers to data exchange between the two via an industrial bus or wireless transmission protocol. Specifically, it can be implemented using Modbus, Profibus, or LoRa communication protocols to ensure the real-time transmission of operating frequency adjustment commands. The program code refers to a set of computer instructions that implement multi-scale time series prediction and operational control algorithms. Specifically, it can be implemented by integrating multilayer perceptrons, cross-scale attention mechanisms, and linear neural network modules, used to map historical operating data to predicted results for future time points.

[0129] Specifically, the sensing components continuously collect temperature and pressure data from the dust collector fan and transmit it to the control module. The control module's processor calls the program code in memory to perform multi-scale decomposition and cross-scale feature fusion on the original time series, generating predictions of future operating conditions. Based on the correlation between the prediction results and the operating frequency, the control module calculates the target frequency and sends it to the frequency converter via the communication interface. The frequency converter converts the electrical signal into a motor drive signal, dynamically adjusting the fan speed. When the dust collection system is in a high-temperature and high-pressure environment, this process can be repeated cyclically to achieve real-time closed-loop control.

[0130] This solution combines multi-scale time-series prediction with a real-time communication mechanism to achieve high-frequency data acquisition and processing under ultra-long ventilation duct conditions, reducing prediction deviations caused by environmental interference and thus improving control response speed. Through the above technical solution, this application solves the problem of response delay in dust removal systems in high-temperature and high-pressure industrial environments. By combining multi-scale prediction with frequency conversion control, the synchronization between fan speed adjustment and operating condition changes is ensured, avoiding the harm to the health of on-site personnel caused by the decline in dust removal efficiency.

[0131] Time series data exhibits different characteristics across different fields, and sequence data collected at different times and in different environments will show different patterns. This is also true for time series data from sensors in dust collector fans, which will show different patterns under different operating conditions. This application focuses on time series data from dust collector fans, first using multilayer perceptron processing to capture the nonlinearity of the data, then extracting high-order features at different granularities through multi-scale decomposition and fusion, and finally, considering the differences in prediction results at different scales, weighted summing of the prediction results at multiple scales to obtain the final prediction result. The prediction results for each operating condition are as follows: Figures 7 to 9 As shown in the figure, the predicted values ​​obtained by the method of this application almost coincide with the true values, compared with the control group. Figures 10 to 12 In comparison, the results of comparing the effects of the three sets of prediction data show that the predicted values ​​of this application have a higher degree of overlap with the actual values, and the prediction effect is better.

[0132] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the terminals, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0133] In the several embodiments provided in this application, it should be understood that the disclosed terminals, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0134] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0135] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0136] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0137] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0138] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0139] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A multi-scale time series prediction method, characterized in that, include: Obtain the original time series; The original time series is decomposed to obtain multiple sub-time series with different particle scales; Based on the order of granularity of each sub-time series, the first two sub-time series are taken as the first and second series respectively. Then, based on the first and second series, QKV features and attention weights corresponding to the QKV features are calculated. The Q features are calculated based on the first series, while the K and V features are calculated based on the features of the second series. Based on the second sequence and the attention weight, feature fusion processing is performed to obtain a fused time series. If the next sub-time series corresponding to the current second sequence is not empty, the first sequence is updated based on the fused time series, and the second sequence is updated based on the next sub-time series. Based on the updated first and second sequences, the latest fused time series is obtained. If the next sub-time series corresponding to the current second sequence is empty, a multi-scale time series is constructed based on the first sub-time series and each fused time series. Time series prediction is performed based on the multi-scale time series to obtain the prediction results.

2. The multi-scale time series prediction method according to claim 1, characterized in that, Also includes: The original time series is preprocessed by performing a nonlinear transformation using a pre-defined multilayer perceptron.

3. The multi-scale time series prediction method according to claim 2, characterized in that, The multilayer perceptron includes: a layer normalization module, a fully connected layer, a GELU activation function, a Dropout calculation module, and a residual connection processing module.

4. The multi-scale time series prediction method according to claim 1, characterized in that, The step of decomposing the original time series to obtain multiple sub-time series at different granular scales includes: The original time series is decomposed by average pooling downsampling to obtain multiple sub-time series with different granular scales.

5. The multi-scale time series prediction method according to claim 1, characterized in that, The feature fusion process based on the second sequence and the attention weights to obtain the fused time series includes: The sum of the second sequence and the attention weights is subjected to layer normalization to obtain a preliminary fusion result; The preliminary fusion result is subjected to high-order feature extraction by a preset feedforward neural network. The extracted high-order features are then subjected to residual connection and layer normalization processing with the preliminary fusion result in sequence to obtain the fused time series.

6. The multi-scale time series prediction method according to claim 1, characterized in that, The time series prediction based on the multi-scale time series yields prediction results including: The multi-scale time series is input into a preset linear neural network, and the prediction result is determined based on the output of the linear neural network.

7. The multi-scale time series prediction method according to claim 6, characterized in that, The determination of the prediction result based on the output of the linear neural network includes: Based on the multi-scale prediction result sequence output by the linear neural network, and combined with a preset learnable weight coefficient sequence, the elements in the multi-scale prediction result sequence are weighted and summed to obtain the final prediction result based on the summation result.

8. A multi-scale time series prediction device, characterized in that, include: The raw sequence acquisition unit is used to acquire the raw time series. A multi-scale sequence decomposition unit is used to decompose the original time series into multiple sub-time series with different particle scales. A cross-scale attention processing unit is used to take the first two sub-time series as the first sequence and the second sequence in order of the granularity of each sub-time series, and then calculate the QKV features and the attention weights corresponding to the QKV features based on the first sequence and the second sequence. The Q features are calculated based on the first sequence, and the K and V features are calculated based on the features of the second sequence. A cross-scale feature fusion unit is used to perform feature fusion processing based on the second sequence and the attention weight to obtain a fused time series. If the next sub-time series corresponding to the current second sequence is not empty, the first sequence is updated based on the fused time series, and the second sequence is updated based on the next sub-time series, so as to obtain the latest fused time series based on the updated first and second sequences. If the next sub-time series corresponding to the current second sequence is empty, a multi-scale time series is constructed based on the first sub-time series and each fused time series. The time series prediction unit is used to perform time series prediction based on the multi-scale time series and obtain the prediction results.

9. A method for controlling the operation of a dust collector fan, characterized in that, include: Collect operating data of the dust collector fan at multiple time points; Using the operating data at the multiple time points as the original time series, the operating data prediction result of the dust removal fan at a preset future time is obtained according to the multi-scale time series prediction method as described in any one of claims 1 to 7. Based on the predicted operating data, and combined with the correlation between the operating data and the operating frequency of the dust collector fan, the target operating frequency corresponding to the predicted operating data is determined. The target operating frequency is transmitted to the frequency converter so that the frequency converter adjusts the fan speed according to the target operating frequency.

10. A dust removal system, characterized in that, include: Dust collector fan, frequency converter, control module and several sensing components; The sensing component is installed in the dust removal fan and is used to collect the operating condition data of the dust removal fan at multiple time points. The control module is communicatively connected to the frequency converter and each of the sensing components, and the signal output terminal of the frequency converter is communicatively connected to the control terminal of the dust removal fan. The control module includes a memory and a processor. The memory is used to store program code, which corresponds to the dust collector fan operation control method as described in claim 9. The processor is used to read and execute the program code to implement the dust collector fan operation control method.

Images