Article: Region-specific multi-scale meteorological forecasting based on data assimilation and reinforcement learning Journal: International Journal of Reasoning-based Intelligent Systems (IJRIS) 2026 Vol.18 No.8 pp.1 - 9 Abstract: Accurate meteorological forecasting is vital for disaster prevention. However, existing approaches often suffer from significant heterogeneity in meteorological data. To address these challenges, this paper introduces a data assimilation method based on particle swarm optimisation and particle filtering to derive assimilated meteorological observation variables. Subsequently, the seasonal-trend decomposition using LOESS is applied to disaggregate meteorological series. The trend component is predicted using a gated recurrent unit model, while the seasonal and residual components are formulated as state variables. This reformulation transforms forecasting problems into the multi-dimensional decision-making task, facilitating the training of a reinforcement learning model to improve forecasting accuracy. Experimental results show that the proposed model reduces the root mean square error by at least 13.93% and 15.21% for forecast lead times of 6 and 24 days, respectively, demonstrating its potential as an effective technical solution for high-precision meteorological forecasting across diverse climatic regions. Inderscience Publishers - linking academia, business and industry through research

Title: Region-specific multi-scale meteorological forecasting based on data assimilation and reinforcement learning

Authors: Zhenhong Sun; Xi Liu; Yicen Liu; Ruohan Li

Addresses: State Grid Sichuan Electric Power Research Institute, Chengdu, 610041, China ' State Grid Sichuan Electric Power Research Institute, Chengdu, 610041, China ' State Grid Sichuan Electric Power Research Institute, Chengdu, 610041, China ' State Grid Sichuan Electric Power Research Institute, Chengdu, 610041, China

Abstract: Accurate meteorological forecasting is vital for disaster prevention. However, existing approaches often suffer from significant heterogeneity in meteorological data. To address these challenges, this paper introduces a data assimilation method based on particle swarm optimisation and particle filtering to derive assimilated meteorological observation variables. Subsequently, the seasonal-trend decomposition using LOESS is applied to disaggregate meteorological series. The trend component is predicted using a gated recurrent unit model, while the seasonal and residual components are formulated as state variables. This reformulation transforms forecasting problems into the multi-dimensional decision-making task, facilitating the training of a reinforcement learning model to improve forecasting accuracy. Experimental results show that the proposed model reduces the root mean square error by at least 13.93% and 15.21% for forecast lead times of 6 and 24 days, respectively, demonstrating its potential as an effective technical solution for high-precision meteorological forecasting across diverse climatic regions.

Keywords: multi-scale meteorological forecasting; data assimilation; reinforcement learning; seasonal-trend decomposition using LOESS; gated recurrent unit.

DOI: 10.1504/IJRIS.2026.151725

International Journal of Reasoning-based Intelligent Systems, 2026 Vol.18 No.8, pp.1 - 9

Received: 04 Sep 2025
Accepted: 08 Dec 2025
Published online: 17 Feb 2026 *

Title: Region-specific multi-scale meteorological forecasting based on data assimilation and reinforcement learning

Keep up-to-date