Weather forecasting plays a critical role in understanding local atmospheric conditions, yet most centralized models require continuous internet access and produce generalized outputs that overlook local variations. This study presents the deployment of a Multilayer Perceptron (MLP) weather forecasting model on a Raspberry Pi 4 as an edge-capable, offline prototype for hyper-local monitoring. The system collected real-time temperature, humidity, pressure, and rainfall data using onboard sensors and generated 24-hour forecasts evaluated throughRM SE,M AE,sM AP E,M ASE, andR2, alongside ▵RM SEagainst an Exponential Smoothing (ETS) baseline and residual quantile analysis (P10-P90). Results showed that the MLP achieved higher accuracy and stability than ETS baseline model, withR2exceeding 0.97 across most targets. Edge deployment analysis demonstrated low-latency performance (<0.05 ms), stable resource utilization, and a 65% reduction in model size through INT8 quantization. Forecast errors remained within sensor tolerance ranges, validating the system’s feasibility for real-time, embedded weather forecasting under offline conditions.
Rusiana et al. (Wed,) studied this question.