Reinforcement Learning to Discover a North-East Monsoon Index for Rainfall Prediction in Thailand

Accurately predicting long-term rainfall is challenging. Global climate indices, such as the El Niño-Southern Oscillation, are standard input features for machine learning. However, a significant gap persists regarding local-scale indices capable of improving predictive accuracy in specific regions of Thailand. This paper introduces a novel North-East monsoon climate index calculated from sea surface temperature to reflect the climatology of the boreal winter monsoon. To optimise the calculated areas used for this index, a Deep Q-Network reinforcement learning agent explores and selects the most effective rectangles based on their correlation with seasonal rainfall. Rainfall stations were classified into 12 distinct clusters to distinguish rainfall patterns between southern and upper Thailand. Experimental results show that incorporating the optimised index into Long Short-Term Memory models significantly improves long-term monthly rainfall prediction skill in most cluster areas. This approach effectively reduces the Root Mean Square Error for 12-month-ahead forecasts.

💡 Research Summary

This paper tackles the persistent challenge of improving long‑term rainfall forecasts for Thailand by introducing a locally‑tailored climate index that captures the dynamics of the North‑East Monsoon (NEM). While global teleconnection indices such as ENSO, PDO, and MJO are routinely employed as predictors in machine‑learning‑based rainfall models, they often fail to represent regional variability that is crucial for Thailand’s heterogeneous climate. To fill this gap, the authors devise a novel NEM index derived from sea‑surface temperature (SST) anomalies. The index is defined as the difference between the mean SSTs over two rectangular oceanic regions (designated A and B). Rather than manually selecting these regions, a Deep Q‑Network (DQN) reinforcement‑learning agent is tasked with discovering the optimal locations and extents of the rectangles.

The reinforcement‑learning formulation treats the current coordinates of the two rectangles as the state. Actions consist of discrete shifts of ±0.5° in latitude and longitude, and, in a second configuration, also include resizing operations (shrink/expand by ±0.5°). Two action spaces are examined: “shift‑only” (8 actions) and “shift‑and‑resize” (16 actions). The reward is a season‑aware objective function that aggregates the squared Pearson correlation between the candidate NEM index and cluster‑averaged rainfall for both the NEM onset (October‑March, primarily affecting southern Thailand) and the NEM retreat (April‑September, influencing northern Thailand). By maximizing this reward, the agent simultaneously seeks strong negative correlations in both seasons, ensuring the index reflects the monsoon’s dual impact.

Training proceeds for 100 000 timesteps with a discount factor of 0.99, an exploration probability of 0.1, and standard DQN hyper‑parameters. The shift‑only configuration achieves a Q‑value of 0.497, a substantial improvement over the baseline rectangular definition taken from prior literature (Q = 0.052). The shift‑and‑resize configuration yields a Q‑value of 0.412, still markedly better than the baseline. Optimised rectangle coordinates are reported for both settings, demonstrating that the agent discovers geographically plausible SST zones that differ from the originally hypothesised areas.

Data sources comprise two complementary gauge networks. The Thailand Meteorological Department (TMD) provides 74 high‑quality stations spanning 1982‑2024, which are used for DQN training and later for LSTM model fitting. The Hydro‑Informatics Institute (HII) contributes 384 stations (2014‑2024) that serve to construct spatial clusters via hierarchical clustering. After normalising monthly rainfall, applying PCA, and computing Euclidean distances, 12 clusters emerge: clusters 1‑4 represent southern Thailand, while clusters 5‑12 cover the northern and central regions. These clusters are the basis for evaluating the NEM index’s correlation with observed precipitation.

For the predictive component, a Long Short‑Term Memory (LSTM) network is employed with a 24‑month input window and a 12‑month forecast horizon. Input features include the conventional global indices (DMI, MEI, PDO, MJO, BSISO, South‑West Monsoon Index, ONI) and, when applicable, the optimised NEM index. Feature selection retains only those predictors whose absolute correlation with the cluster‑average rainfall exceeds 0.6, ensuring that the NEM index is incorporated only where it demonstrably contributes. Two training folds are used: (1) 1982‑2019 for training, 2020 for validation, 2021 for testing; and (2) 1982‑2022 for training, 2023 for validation, 2024 for testing. Model performance is measured by Root Mean Square Error (RMSE) in mm month⁻¹.

Results show consistent gains when the NEM index is added. In southern clusters 1, 2, 4, RMSE drops from 99.79 → 94.54, 82.61 → 77.05, and 130.02 → 121.48 mm month⁻¹ respectively. Northern clusters also benefit: cluster 6 improves from 57.27 → 53.94, cluster 9 from 56.11 → 55.44, and cluster 12 from 59.24 → 52.38 mm month⁻¹. Clusters where the NEM index’s correlation fell below the 0.6 threshold (3, 7, 10, 11) were excluded, underscoring the index’s region‑specific relevance. Overall, the inclusion of the optimised NEM index reduces 12‑month‑ahead RMSE by roughly 5‑10 % across most of Thailand, confirming its utility for water‑resource planning and drought/flood risk management.

The authors acknowledge several limitations. The DQN’s discrete action space may restrict finer‑grained area optimisation; a continuous‑action algorithm such as DDPG or PPO could yield more precise SST region definitions. The objective function relies solely on linear correlation, potentially overlooking non‑linear relationships between SST and rainfall. Moreover, only LSTM is evaluated; future work should benchmark the NEM index against more sophisticated spatio‑temporal architectures (Convolution‑LSTM, Graph Neural Networks, Neural Operators, state‑space models). Finally, assessing the index’s robustness under climate‑change scenarios (e.g., RCP 8.5) would be essential for long‑term planning.

In summary, this study presents a novel, data‑driven framework that couples reinforcement learning with climate‑index construction, delivering a region‑specific NEM index that materially enhances long‑term rainfall forecasts for Thailand. The methodology is readily extensible to other monsoon‑dominated regions and exemplifies the productive synergy between climate science and modern reinforcement‑learning techniques.