Accurate wind power forecasting is essential for the effective integration of wind energy into power grids. Yet, the inherent variability of wind and the intricate interplay of meteorological factors make prediction a challenging task. This study introduces a novel short-term wind power forecasting method, improving the traditional convolutional neural network and long short-term memory (CNN-LSTM) model through two significant innovations. First, we introduce a parallel convolutional architecture that employs both 1dimensional (1D) and 2-dimensional (2D) convolutions to simultaneously capture temporal patterns and inter-variable relationships in wind power data. This structure, inspired by Explainable-CNNs, enables more comprehensive feature extraction. Second, we integrate an attention mechanism that dynamically weights the importance of different input features and time steps, improving both forecast accuracy and model interpretability. The proposed model is evaluated using data from two wind farms in Croatia, comparing its performance against benchmark models including standard CNN-LSTM, LSTM, and gated recurrent unit (GRU) networks. Results demonstrate that our enhanced CNN-LSTM model achieves superior forecasting accuracy, with improvements in Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of 15% and 12% respectively, compared to the best-performing benchmark. Furthermore, the attention mechanism provides valuable insights into the relative importance of different features over time, offering a new level of interpretability in wind power forecasting models. This work contributes to the advancement of accurate and explainable wind power prediction, supporting more efficient renewable energy integration and grid management.