Evaluation of Transfer Learning Using RNN and LSTM for Room Temperature Prediction Considering Occupancy Variation

Enhancing room temperature prediction performance is beneficial for optimizing air conditioning control and supporting other applications. While previous studies have explored the effective use of prediction models across different buildings through transfer learning, performance under varying occupancy remains unclear. In this study, we evaluated the effectiveness of transfer learning by applying a model trained on data from an office with constant occupancy in Nagoya to a variable occupancy scenario in Hokkaido, adding randomness. The results showed that transfer learning significantly improved prediction accuracy with varying occupancy. The greatest improvements were in short-term predictions up to 1 minute ahead, with a 72% error reduction, and in long-term predictions up to 24 hours ahead, with a 47% reduction. In the future, We plan to further enhance accuracy by refining model and adjusting data used.