BRD-LLAMA based Bike Rental Demand Prediction

Pre-trained large language models (LLMs) have been widely applied in various tasks. They have been proven advantageous in transfer learning and automatically capturing complex features. However, LLMs are trained with discrete vocabularies and cannot directly process continuous numerical time series data in tasks such as bike rental demand prediction. In addition, the large parameter scale of LLM proposes high demands on hardware resources during training. Therefore, we proposed a bike rental demand prediction method BRD-LLAMA in this paper. Firstly, we introduce a learnable sequence embedding method called Series Embedding, which aligns time series data with natural language. Secondly, we introduced Prompt-Tuning to combine domain expert knowledge and task instructions to guide the model's inference process. Finally, Low Rank Adaptation (LoRA) tuning is adopted to improve the training efficiency. Experiments on the London Bike Rental Demand dataset demonstrate that the proposed BRD-LLAMA has better prediction performance.