Soft Periodic Convolutional Recurrent Network for Spatiotemporal Climate Forecast

Ekasit Phermphoonphiphat,Ken-Ichi Fukui,Takashi Morita,Masayuki Numao,Tomohiko Tomita

doi:10.3390/app11209728

Ekasit Phermphoonphiphat, Ken-Ichi Fukui + Show 3 more

Open Access

https://doi.org/10.3390/app11209728

Copy DOI

Journal: Applied sciences	Publication Date: Oct 18, 2021
License type: CC BY 4.0

Affiliation: Osaka University, Kumamoto University

Abstract

Many machine-learning applications and methods are emerging to solve problems associated with spatiotemporal climate forecasting; however, a prediction algorithm that considers only short-range sequential information may not be adequate to deal with periodic patterns such as seasonality. In this paper, we adopt a Periodic Convolutional Recurrent Network (Periodic-CRN) model to employ the periodicity component in our proposals of the periodic representation dictionary (PRD). Phase shifts and non-stationarity of periodicity are the key components in the model to support. Specifically, we propose a Soft Periodic-CRN (SP-CRN) with three proposals of utilizing periodicity components: nearby-time (PRD-1), periodic-depth (PRD-2), and periodic-depth differencing (PRD-3) representation to improve climate forecasting accuracy. We experimented on geopotential height at 300 hPa (ZH300) and sea surface temperature (SST) datasets of ERA-Interim. The results showed the superiority of PRD-1 plus or minus one month of a prior cycle to capture the phase shift. In addition, PRD-3 considered only the depth of one differencing periodic cycle (i.e., the previous year) can significantly improve the prediction accuracy of ZH300 and SST. The mixed method of PRD-1, and PRD-3 (SP-CRN-1+3) showed a competitive or slight improvement over their base models. By adding the metadata component to indicate the month with one-hot encoding to SP-CRN-1+3, the prediction result was a drastic improvement. The results showed that the proposed method could learn four years of periodicity from the data, which may relate to the El Niño–Southern Oscillation (ENSO) cycle.

Highlights

Increasing climate remote sensing, including reanalysis data that combine numerical simulations with observations, generates a massive amount of data
The results show that when increasing the length of sequential input, the root mean square error (RMSE) of all models is significantly decreasing
The RMSE results tend to be converged after the sequence length of five or six months, indicating that the five to six months prior knowledge of ZH300 is essential for one step ahead of ZH300 prediction, which is consistent with the sea surface temperature (SST) memory is about six months [52]

Summary

Introduction

Increasing climate remote sensing (e.g., weather satellite, AMeDAS), including reanalysis data that combine numerical simulations with observations, generates a massive amount of data. It consumes a lot of computational resources to make a forecast by using numerical weather prediction (NWP) models [1], and requires climatology knowledge to build a model, unlike machine learning (ML). Time-series forecasting models [8] have emerged to forecast climate variables by learning from the historical data. Recent works in ML [12,13,14] have proven that LSTM is beneficial for solving time-series climate forecasting problems

Methods

Results

Discussion

Conclusion