Existing electricity theft detection methods primarily focus on high-loss feeders, where anomalies in the daily average line-loss rate are pronounced. However, potential short-term and intermittent electricity theft on non-high-loss feeders is easily masked by load fluctuations, metering errors, and changes in operating conditions, making such theft difficult to detect effectively. To address this issue, this paper proposes a method for electricity theft detection on non-high-loss feeders based on a mutual-information-difference dynamic graph sequence and adjacency features. First, at an hourly resolution, a sliding window is employed to calculate the mutual information between users′ electricity consumption and line-loss electricity, thereby constructing a mutual-information-difference dynamic graph sequence. A window-wise adaptive threshold is introduced to capture the nonlinear dependency between users′ electricity usage and line loss as it evolves over time, avoiding graph-structure instability caused by a fixed threshold. Second, a feature extraction method combining a graph autoencoder and a dynamic weight balancing mechanism is developed. Based on a cross-window adjacency reconstruction task, it captures the temporal evolution of connection patterns within the population network and extracts adjacency features that reflect the robustness of users′ dependency relationships with the group. Meanwhile, a dynamically weighted loss function is applied to mitigate training bias caused by the imbalance between positive and negative samples and to prevent degeneration in graph reconstruction learning. Finally, principal component analysis is used to reduce and select the embedded features, and K-Means++ clustering is applied for unsupervised anomaly identification. Simulation studies and experiments on real-world data demonstrate that the proposed method can effectively detect electricity-theft users in non-high-loss scenarios, verifying its feasibility and effectiveness. From the perspective of group relationship evolution, this method characterizes the dynamic dependencies between users′ electricity consumption and line loss, transforms electricity theft detection into a graph anomaly detection problem, and provides an effective and practically applicable technical solution for electricity theft detection on medium-voltage feeders.