Data-prediction method with multi-time-scale, monitoring system using the same, and non-transitory readable medium

Abstract

A multi-timescale data prediction method, monitoring system, and non-transitory readable recording medium relate to a multi-timescale monitoring system, which includes a storage unit and a computing engine. The storage unit stores a prediction model. This prediction model is configured to perform multi-timescale calculations on one-dimensional feature data to obtain prediction results. The computing engine is coupled to the storage unit and is configured to load and execute this prediction model.

Claims

1. A multi-timescale monitoring system, comprising: A storage unit stores a prediction model, which is configured to perform multi-timescale calculations on one-dimensional feature data to obtain a prediction result. And a computing engine, coupled to the storage unit, configured to load and execute the prediction model; The prediction model includes: a data conversion module, configured to convert the one-dimensional feature data into a two-dimensional image set; a feature extraction module, coupled to the data conversion module, configured to extract features from the two-dimensional image set to obtain a time-step sequence of the one-dimensional feature data; a first sub-neural network system, coupled to the feature extraction module, configured to obtain a first encoded feature based on the time-step sequence; a second sub-neural network system, configured to obtain a second encoded feature based on the one-dimensional feature data; a gating fusion module, coupled to the first and second sub-neural networks, configured to perform feature fusion and weighting on the first and second encoded features using a gating mechanism to obtain a final feature vector; and an output module, coupled to the gating fusion module, configured to map the final feature vector onto the dimension of future prediction time to obtain a prediction result.

2. The multi-timescale monitoring system as described in claim 1, wherein the one-dimensional feature data is an energy consumption feature, and the computing engine is configured to further initiate a load shedding mechanism and / or perform power generation resource scheduling based on the prediction results.

3. The multi-timescale monitoring system as described in claim 1, wherein the one-dimensional feature data has complex feature attributes, the two-dimensional image set is a multi-channel input tensor, and the data conversion module includes: A normalization unit is configured to normalize the one-dimensional feature data of each feature attribute; A transformation unit, coupled to the normalization unit, is configured to transform the normalized one-dimensional feature data into angles in a polar coordinate system to obtain polar coordinate data for each feature attribute; a generation unit, coupled to the transformation unit, is configured to generate Gram matrices for each feature attribute based on the polar coordinate data for each feature attribute; and a combination unit, coupled to the generation unit, is configured to combine the Gram matrices of the complex feature attributes into the multi-channel input tensor.

4. The multi-timescale monitoring system as described in claim 3, wherein the one-dimensional feature data is time series data of the complex feature attribute, the time series data of each feature attribute is composed of the values ​​of the feature attribute at complex time points, and the prediction model further includes: A preprocessing module, coupled to the data conversion module and the second sub-neural network system, is configured to assemble the raw data of the complex feature attributes into a feature matrix, standardize each column of the feature matrix, and extract the value of the complex time point of each feature attribute from the standardized feature matrix to obtain the time series data of each feature attribute.

5. The multi-timescale monitoring system as described in claim 1, wherein the one-dimensional feature data is time-series data with a single feature attribute.

6. The multi-timescale monitoring system as described in claim 1, wherein the feature extraction module comprises: A convolutional layer, coupled to the data transformation module, has complex convolutional kernels of different sizes; A fully connected layer is coupled to the first sub-neural network system; And a pooling layer, coupled between the convolutional layer and the fully connected layer.

7. The multi-timescale monitoring system as described in claim 1, wherein the first sub-neural network system and the second sub-neural network system have the same computational structure.

8. The multi-timescale monitoring system as described in claim 7, wherein the computational structure includes: A position encoding module, configured to position-encode input data, wherein the input data is the time-step sequence or the one-dimensional feature data; an encoder-decoder structure coupled to the position encoding module, configured to use a self-attention mechanism to perform converter encoding on the output data of the position encoding module; and a time-aware module coupled between the encoder-decoder structure and the gated fusion module, configured to use a time-aware attention mechanism to perform weighted multiplication of the output data of the encoder-decoder structure with corresponding time features, wherein the output data of the time-aware module corresponds to the input data as the first encoded feature or the second encoded feature.

9. The multi-timescale monitoring system as described in claim 1, wherein the gating fusion module uses a feedforward neural network to perform the feature fusion and the weighting.

10. A multi-timescale data prediction method, comprising: One-dimensional feature data is converted into a two-dimensional image set; features are extracted from the two-dimensional image set to obtain a time step sequence of the one-dimensional feature data; a first encoded feature is obtained based on the time step sequence; a second encoded feature is obtained based on the one-dimensional feature data; a gating mechanism is used to fuse and weight the first encoded feature and the second encoded feature to obtain a final feature vector; and the final feature vector is mapped to the dimension of future prediction time to obtain a prediction result.

11. The multi-timescale data prediction method as described in claim 10, wherein the one-dimensional feature data is an energy consumption feature, and the data prediction method further includes: Based on the forecast results, a load release mechanism will be activated and / or power generation resources will be dispatched.

12. The multi-timescale data prediction method as described in claim 10, wherein the one-dimensional feature data has complex feature attributes, the two-dimensional image set is a multi-channel input tensor, and the transformation step of the two-dimensional image set includes: The one-dimensional feature data of each feature attribute is normalized; the normalized one-dimensional feature data is converted into angles in the polar coordinate system to obtain the polar coordinate data of each feature attribute; the Gram matrix of each feature attribute is generated based on the polar coordinate data of each feature attribute; and the Gram matrix of the complex feature attributes is combined into the multi-channel input tensor.

13. The multi-timescale data prediction method as described in claim 12, wherein the one-dimensional feature data is time series data of the complex feature attribute, and the time series data of each feature attribute is composed of the values ​​of the feature attribute at complex time points, and the data prediction method further includes: Receive the raw data of the complex number's characteristic attribute; The original data of the complex feature attributes are assembled into a feature matrix; Each column of the feature matrix is ​​standardized; and the complex time point value of each feature attribute is extracted from the standardized feature matrix to obtain the time series data of each feature attribute.

14. The multi-timescale data prediction method as described in claim 10, wherein the one-dimensional feature data is time series data composed of the values ​​of a complex number of time points of a single feature attribute.

15. The multi-timescale data prediction method as described in claim 10, wherein the feature extraction steps include: Multi-scale features are extracted from this set of two-dimensional images using complex convolution kernels and activation functions, wherein the sizes of the complex convolution kernels are different from each other; The multi-scale features are sampled to obtain sampled features; and the sampled features are integrated to obtain the time step sequence, wherein the time step sequence has a given tensor.

16. The multi-timescale data prediction method as described in claim 10, wherein the first encoded feature and the second encoded feature are obtained by processing the time step sequence and the one-dimensional feature data respectively using the same computational structure.

17. The multi-timescale data prediction method as described in claim 16, wherein the step of obtaining the first encoded feature includes: The time step sequence is position-encoded to obtain a time step sequence with position information; A self-attention mechanism is used to encode the time step sequence containing the location information to obtain the first output feature; and a time-aware attention mechanism is used to multiply the first output feature by the weights of the corresponding time features to obtain the first encoded feature.

18. The multi-timescale data prediction method as described in claim 16, wherein the step of obtaining the second encoded feature includes: The one-dimensional feature data is positionally encoded to obtain the one-dimensional feature data with positional information. A self-attention mechanism is used to encode the one-dimensional feature data with the location information to obtain the second output feature; and a time-aware attention mechanism is used to multiply the second output feature by the weights of the corresponding time features to obtain the second encoded feature.

19. The multi-timescale data prediction method as described in claim 10, wherein the step of obtaining the final feature vector includes: A feedforward neural network is used to perform feature fusion and weighting on the first encoded feature and the second encoded feature.

20. A non-transitory readable recording medium storing at least one program, which, when loaded and executed by a computer, enables a multi-timescale data prediction method as described in any one of claims 10 to 19.

Images