FeatureFuser-LLM: Multi-Scale Feature Fusion with Adaptive Positional Encoding for LLM-Based Time Series Forecasting

Long-term time series forecasting (LTSF) plays a crucial role in many modern intelligent systems. With the boost of large language models (LLMs), LTSF modeling can leverage abundant knowledge from multiple domains, enhancing the universality of these models. In this work, we propose FeatureFuser-LLM, a framework combining multi-scale CNNs with frozen LLMs for time series forecast. Cross-channel feature representation is obtained by performing convolutions at the macro-, meso-, and micro-levels, followed by an adaptive positional encoding tensor to capture temporal dependencies. The recently released DeepSeek-R1 distilled model series are selected as the frozen LLM backbone of our method, which retains excellent numerical reasoning capabilities while maintains a relatively small parameter size. Experiments across datasets from different domains under two training set ratios were conducted. FeatureFuser-LLM outperforms other baselines, especially in few-shot scenarios.