In Sri Lanka, grave crimes such as homicide, robbery, and violent assaults were compromising public safety and social stability. Understanding their spatio-temporal patterns was important for effective law enforcement, resource allocation, and crime prevention. This study aimed to forecast weekly crime counts at the police station level by applying advanced spatio-temporal models. A dataset of 151,032 crime records from 2019 to 2023, obtained from the Police Criminal Records Division (CRD) in Sri Lanka. The dataset included spatial and temporal variables along with crime incident identifiers. Descriptive analysis revealed evident spatial and temporal crime patterns with strong spatial concentration and clustering in urban areas, particularly in the Western province, and crime peaks that coincide with public holidays. Property and financial offenses also substantially increased in 2023. These patterns motivated the use of advanced spatio-temporal models to capture the underlying dynamics more effectively. Several models were tested, including the spatio-temporal Log-Gaussian Cox Process (ST-LGCP), a zero-inflated negative binomial generalized linear mixed model (ZINB-GLMM), and spatio-temporal deep learning models such as residual networks. Among them, the ST-LGCP performed the best, with the lowest test RMSE of 1.4439. The study found that month, week of the month, province, and police station were significant variables. These findings highlight that, for this dataset, spatio-temporal modeling, especially the ST-LGCP, performs well in forecasting weekly crime counts in Sri Lanka, supporting strategic law enforcement planning and decision-making.