Electricity theft seriously threatens the secure and stable operation of power systems. With the continuous upgrading of metering systems, the availability of multi-source data has been significantly improved, providing favorable conditions for electricity theft detection research. However, existing methods still suffer from the limitations of single-dimensional data utilization and the difficulty of effectively capturing complex temporal dependencies. To address these challenges, this paper proposes a parallel convolutional neural network self-attention network (PCSA-Net) for electricity theft detection, which aims to exploit the complex coupling relationships among heterogeneous electricity, gas and water consumption data. In the proposed framework, multi-dimensional heterogeneous time-series data are first normalized to ensure the consistency and reliability of model inputs. PCSA-Net is composed of parallel convolutional neural networks for independent feature extraction from different energy sequences, convolutional layers for cross-channel feature fusion, and a self-attention mechanism for temporal dependency modeling and feature weighting. To alleviate the impact of class imbalance in electricity theft detection, a weighted binary cross-entropy loss function is employed during model training. Extensive experiments conducted on multiple datasets with different anomaly ratios demonstrate that PCSA-Net consistently outperforms traditional and deep learning baseline methods in terms of area under curve (AUC) and weighted F1-score (WF1), confirming its effectiveness in complex multi-source environments.