<p>Constrained by the spatio-temporal bias of traditional manual parking ticket data, it is difficult to understand the distribution of future parking violations demand and the role of influencing factors, parking violations management has primarily relied on post facto approaches. There is a lack of dynamic identification and proactive management tools for high-incidence parking violations areas. Hence, it is crucial to develop an on-street short-term prediction model for parking violations. To achieve this, we utilized time-series data from 80,572 electronic police parking violations captured over a six-month period for the first time. We constructed a variable system incorporating built environment factors (land use, transportation supply, and road design), natural environment factors, and historical characteristics of parking violations demand. Considering also the spatio-temporal heterogeneous effects of built environment on parking violations, we constructed a multiscale geographically and temporally weighted regression model (MGTWR). Further taking spatio-temporal coefficients of MGTWR as inputs to the built environment, we developed a new deep learning framework that integrates MGTWR ResNet, GAT, and attention LSTM (called “MGTWR-RGA”). We used the Dadukou district of Chongqing as an empirical subject, and the results show that the prediction error of the number of parking violations is within 1 on average. We compared the proposed method with the baseline method, and the results show that MGTWR-RGA outperforms the baseline method, and finally, the ablation test verifies that the model is not redundant. These results will support the management of parking violations.</p>