<![CDATA[Hydrometeorological disasters dominate annual disaster occurrences in Indonesia, yet local rainfall prediction remains challenging due to atmospheric complexity and limited resolution of numerical weather models. Server-based forecasting systems further depend on high-performance computing infrastructure and stable network connectivity that are not always available in the field. This study develops an hourly rainfall intensity classification model based on One-Dimensional Convolutional Neural Network (1D-CNN) deployable on an ESP32-S3 microcontroller as a proof-of-concept inference component for rainfall early warning systems. The model uses nine meteorological features arranged in an 18×9 sliding window derived from observational data from the BMKG Soekarno-Hatta Meteorological Station AWS from 2018 to 2025. Logarithmic class weighting and Precision-Recall curve threshold calibration were applied to address extreme class imbalance in hourly resolution data. The model was compared against five baseline models under identical configurations. Threshold calibration increased K2 recall from 0.037 to 0.236 and improved Macro-F1 from 0.461 to 0.530, outperforming all baseline models in terms of Macro-F1. Post-training quantization INT8 reduced model size from 185.8 KB to 64.8 KB with 99.02% decision agreement against the Float32 model. On-device inference on ESP32-S3 achieved a total latency of 13.873 ms with 36.5 KB tensor arena and 234.7 KB free heap, confirming real-time operation without dependence on servers or internet connectivity.]]>
