Lena-TRNN: Exploring energy flow for time series prediction.

We focus on exploring the inherent energy flow for time series prediction in this paper, i.e., we consider the inherent energy of time series data as a sequence measuring properties such as fluctuations, oscillations, and trends. Distinctive with main-stream methods that adopt complex architecture to capture the presentative data relation, this brand-new perspective allows us to better differentiate the underlying distribution of the time-series data. Concretely, we design a novel decoder-free architecture, Latent-energy-aware Transformer Recurrent Neural Network (Lena-TRNN), for multivariate time series forecasting and imputation. The new network tends to assign low energy scores to the samples belonging to the in-distribution dataset, and high energy scores otherwise, which is in accordance with the law of natural change. Predicted samples can be readily obtained by iterating gradient-based optimization along the direction of energy minimization to explicitly learn the inherent energy flow. With the energy optimization modelling, our proposed method exhibits superior performance to many other competitive methods and attains state-of-the-art performance in many benchmark time series forecasting and imputation tasks. The code is available at https://github.com/PengleiGao/Lena-TRNN.

Duke Scholars

Author Kaizhu Huang DKU Faculty