CN-122020067-A - Power grid multivariable time sequence prediction model and method

CN122020067ACN 122020067 ACN122020067 ACN 122020067ACN-122020067-A

Abstract

The invention relates to the technical field of time sequence prediction and deep learning, in particular to a power grid multivariable time sequence prediction model and a method. The time sequence prediction model comprises a time embedding network, a period gating network and a multi-layer perceptron which are sequentially connected, wherein the time embedding network generates time embedding related to the time step length of input data in the form of a multi-variable time sequence based on multi-period time indexes, the period gating network performs normalization processing on the input data in a time dimension and then maps to obtain weights corresponding to all periods used in the time embedding network, the weights and the time embedding of all periods are fused based on a time query mechanism to obtain modulation characteristics, and the multi-layer perceptron processes the dimension superposition results of the input data which are preferentially represented by the modulation characteristics and channels to obtain a time sequence prediction result. The method solves the problems of single periodic modeling mode, channel modeling staticization and high calculation complexity of the multivariable time sequence prediction method in the prior art.

Inventors

SUN RENHAO
FENG JUN
WANG YUJIAN

Assignees

数据空间研究院

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. A training method of a power grid multivariable time sequence prediction model is characterized by comprising the following steps of: Firstly, constructing a time sequence prediction model, which predicts the next multi-variable time sequence based on the existing multi-variable time sequence, wherein the two multi-variable time sequence data structures are identical; The time sequence prediction model comprises a time embedded network, a period gating network and a multi-layer perceptron which are sequentially connected; The time embedding network generates time embedding related to the time step of the input data in the form of a multi-variable time series based on the multi-period time index; the period gating network performs normalization processing on input data in a time dimension to obtain a sample level statistical representation g, and then maps the g to obtain weights corresponding to each period used in the time embedding network Fusing the weight of each period based on the time inquiry mechanism And time embedding Obtaining modulation characteristics ; Modulation features for multi-layer perceptron And processing the dimension superposition result of the input data which is preferentially expressed by the channel to obtain a time sequence prediction result.
2. The training method of the power grid multivariable time series prediction model according to claim 1, wherein the input of the time embedding network is input data in the form of a multivariable time series and the set K periods; Time embedded network construction of index sequences for each period And cycle time feature Then the cycle time is characterized Mapping is time embedding, r is time step index, and step length is 1; ; Wherein, the Is that Is the index value of the r-th index, Mod represents the modulo operation, which is the number of unit times of the kth period.
3. The training method of a multivariate time series prediction model of a power grid according to claim 1, wherein the period gating network fuses weights of each period based on a time inquiry mechanism And time embedding Obtaining modulation characteristics The method comprises the steps of firstly embedding time of each period by combining the period weight Weighting and fusing to obtain time inquiry matrix Q, and preferentially expressing the channel of Q Acquiring channel weight by aggregation and activation in time dimension, and preferentially representing the channel weight and the channel of input data Performing channel-by-channel multiplication to obtain modulation characteristics 。
4. A method of training a multivariate time series prediction model of a power grid as claimed in claim 3 wherein the period gating network incorporates period weights for time embedding of each period The weighted fusion is carried out by , Representing a feature stitching operation.
5. A method for training a multivariate time series prediction model of a power grid according to claim 3, wherein the activation function is a function of The function is activated by firstly carrying out channel-level weighting on the aggregation result and then activating.
6. The method for training a multivariate time series prediction model of a power grid according to claim 1, wherein the cycle-gated network sample level statistical representation g is activated after double-layer mapping to obtain the statistical representation comprising the weights of each cycle Periodic weight vector of (a) 。
7. The method of training a multivariate time series prediction model of a power grid of claim 6, wherein the sample level statistical representation g in the period gating network is activated by a nonlinear activation function after a first mapping and by a Sigmoid activation function after a second mapping.
8. The training method of the power grid multivariable time sequence prediction model according to claim 1, wherein the time sequence prediction model further comprises a preprocessing module, the preprocessing module is used for normalizing the input same batch of multivariable time sequences, and the input data of the time embedded network is the multivariable time sequences after the normalization.
9. A method for predicting the power grid multivariable time series by adopting the training method of the power grid multivariable time series prediction model as claimed in any one of claims 1-8, which is characterized in that the training method of the power grid multivariable time series prediction model as claimed in any one of claims 1-8 is adopted to train a time series prediction model, then the current data is collected to construct a data multivariable time series with a specified length, and the data multivariable time series is input into the time series prediction model to generate the next time series as a prediction result.
10. A system comprising a memory and a processor, the memory having a computer program stored therein, the processor being coupled to the memory, the processor being configured to execute the computer program to implement the grid multi-variable time series prediction method of claim 9.

Description

Power grid multivariable time sequence prediction model and method Technical Field The invention relates to the technical field of time sequence prediction and deep learning, in particular to a power grid multivariable time sequence prediction model and a method. Background The existing multivariable time sequence prediction method is mainly based on a cyclic neural network, a time sequence convolution network or a transducer architecture, and realizes prediction by modeling a dependency relationship in a time dimension or a feature dimension. However, most of the existing methods only implicitly learn the time periodicity, and are difficult to explicitly describe multi-scale periods (such as daily periods and Zhou Zhouqi), and the dependency relationship among channels is usually modeled by fixed parameters or static attention, so that the importance of the channels is difficult to dynamically adjust according to the time context. In addition, some self-attention mechanism-based methods have large calculation and storage overhead in a long sequence scene, and are not beneficial to engineering deployment. Under the current big data trend, the application of the multivariate time sequence prediction is more and more popular, especially in the power grid field, the variables are more, the situation is complex, and the multivariate time sequence prediction has higher requirements on fault prediction, so that the current multivariate time sequence prediction method is difficult to meet the requirements. Disclosure of Invention In order to overcome the defects of single periodic modeling mode, channel modeling staticization and high computational complexity of the multivariate time sequence prediction method in the prior art, the invention provides a multivariate time sequence prediction model for a power grid, which can explicitly describe multicycle time information and dynamically modulate channel weights by utilizing the time information to conduct efficient time sequence prediction. The invention provides a training method of a power grid multivariable time sequence prediction model, which comprises the steps of firstly constructing a time sequence prediction model, predicting the next multivariable time sequence based on the existing multivariable time sequence, wherein the two multivariable time sequence data structures are the same; The time sequence prediction model comprises a time embedded network, a period gating network and a multi-layer perceptron which are sequentially connected; The time embedding network generates time embedding related to the time step of the input data in the form of a multi-variable time series based on the multi-period time index; the period gating network performs normalization processing on input data in a time dimension to obtain a sample level statistical representation g, and then maps the g to obtain weights corresponding to each period used in the time embedding network Fusing the weight of each period based on the time inquiry mechanismAnd time embeddingObtaining modulation characteristics; Modulation features for multi-layer perceptronAnd processing the dimension superposition result of the input data which is preferentially expressed by the channel to obtain a time sequence prediction result. Preferably, the input of the time embedding network is input data in the form of a multi-variable time sequence and K periods are set; Time embedded network construction of index sequences for each period And cycle time featureThen the cycle time is characterizedMapping is time embedding, r is time step index, and step length is 1;; ; ; Wherein, the Is thatIs the index value of the r-th index,Mod represents the modulo operation, which is the number of unit times of the kth period. Preferably, the period gating network fuses the weights of each period based on a time inquiry mechanismAnd time embeddingObtaining modulation characteristicsThe method comprises the steps of firstly embedding time of each period by combining the period weightWeighting and fusing to obtain time inquiry matrix Q, and preferentially expressing the channel of QAcquiring channel weight by aggregation and activation in time dimension, and preferentially representing the channel weight and the channel of input dataPerforming channel-by-channel multiplication to obtain modulation characteristics。 Preferably, the period gating network incorporates period weights for time embedding of each periodThe weighted fusion is carried out by,Representing a feature stitching operation. Preferably, the activation function employsThe function is activated by firstly carrying out channel-level weighting on the aggregation result and then activating. Preferably, the period gating network sample level statistical representation g is activated after double-layer mapping to obtain the sample level statistical representation containing each period weightPeriodic weight vector of (a)。 Preferably, the sample level statistical r