DHRUVILL KALATHIYA

Forecasting framework on large-scale 15-minute heat pump datasets (~3600 units, 90+ types). Benchmarked statistical (Regression), machine learning (XGBoost), and deep learning (LSTM, GRU, Transformer) models. Zero shot Transfer learning approach enabling models trained on one group of devices to generalize to unseen units. Especially effective for newly installed or data-scarce heat pumps, where historical data is limited.

Data and signal analysis of 15-minute time series from heat pump sensors, identifying operational patterns and anomalies. Feature selection and input pattern design to build generalizable representations for forecasting tasks. Clustering of heat pump profiles to group devices with similar behavior and improve model generalization.

data science