Aiming at the poor noise resistance of the exponential smoothing method in short-term load forecasting and the multiple seasonality of power load data, a hybrid forecasting model combining adaptive filters, exponential smoothing, and dilated recurrent neural networks with attention mechanisms was proposed. First, the adaptive spectral module was employed to remove noise from the frequency domain perspective. Subsequently, exponential smoothing was utilized to decompose the load data, remove seasonality, normalize the data, and construct the input vector. Finally, the dilated recurrent neural network combined with the self-attention mechanism dynamically allocates weights based on the components of the current input vector and captured the long-and short-term dependencies in the load data. Additionally, the model learned the parameters of exponential smoothing to enhance the flexibility and accuracy of the forecast. Experimental results demonstrate that the proposed model exhibits superior noise resistance and prediction accuracy.