Malicious users can exploit vulnerabilities of numerous hardware and software components that are integrated into smart grids to launch various physical/cyber attacks for stealing electricity. This causes substantial economic losses and significant security risks. The mainstream electricity theft detection techniques are deep learning-based approaches. However, they present difficulties in finding reliable long-range dependencies from long-term electricity consumption time series, since intricate consumption patterns obscure the temporal dependencies. To solve these problems, we propose an electricity theft detection approach based on the series-wise auto-correlation mechanism, called the ETD-SAC detector. It can progressively decompose intricate consumption patterns throughout the whole detection process and aggregate the dependencies at the sub-series level based on the series-wise auto-correlation mechanism. Experiment results show that the ETD-SAC detector outperforms the state-of-the-art methods in terms of accuracy, false negative rate, and false positive rate.