Data-driven rainfall models normally need long local records, which many stations lack. This study asks whether a transformer pre-trained on data-rich coastal Mumbai can cut both the local data and the computation needed to forecast daily rainfall at Pune, a rain-shadow city 120 km inland. RainFormer, a transformer encoder-decoder, first trained on Pune’s full record, gave a reference test RMSE of 12.522 mm. It was then pre-trained on Mumbai, and only its last decoder block and output head, 52,481 of its 194,177 parameters, were fine-tuned on 10 to 100% of Pune’s record, against models trained from scratch on identical data. Transfer beat the scratch model at every share. With 10% of the local record its RMSE was within 0.43% of the reference, it equalled the reference at 50% and bettered it at 60% (12.441 mm), so full-data accuracy was reached with 40% less local data while training 73% fewer parameters. Beyond that point the advantage narrowed and the transfer RMSE rose slightly, to 12.479 mm at 100%, indicating over-fitting of the fine-tuned layers to Pune’s historical regime. Cross-city transfer can therefore replace much of a station’s record at a fraction of the training cost.

Transfer learning, Transformer, Rainfall forecasting, Fine-tuning, Computational efficiency.