Abstract: With the high proportion of renewable energy integration and the deepening of electricity marketization, higher requirements have been placed on the accuracy and stability of short-term power prediction. Traditional time series prediction models have limitations in handling the nonlinear, non-stationary, and complex temporal dependencies of power data. To address this, this study proposes a deep learning prediction model that integrates adaptive filtering and an attention mechanism. The method first preprocesses the original power series using an adaptive filtering algorithm (such as LMS or RLS) to smooth noise and extract underlying trends. Then, a core prediction model is constructed based on a Long Short-Term Memory (LSTM) encoder-decoder framework embedded with an attention mechanism, enabling the model to dynamically focus on information from key historical time steps. The experimental results demonstrate that the proposed method achieves satisfactory performance in terms of both prediction accuracy and robustness, thereby offering an effective solution for short-term power prediction.