Safety compliance in construction environments is crucial to minimizing accidents and protecting workers. Most of the traditional models of safety rely on a simple form of learning, and this type may not be robust enough to meet the level of complexity and variability of construction sites. This research, therefore, presents a complete method that uses deep learning algorithms to make accurate predictions on the safety compliance at construction sites. We adopt a multi-model approach by combining several machine learning techniques: Logistic Regression, Random Forest (RF), K-Nearest Neighbors (KNNs), Support Vector Machine (SVM), and XGBoost, as well as more powerful deep learning models: Group Method of Data Handling (GMDH), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Fully Connected Neural Networks (FCNN).The model performance has been optimized with the extensive preprocessing procedure including filling missing and null values, correction of outliers, feature engineering, and data standardization. Model training and testing was carried out on different data sets. Metrics that check the model’s performance are accuracy, precision, recall, F1-score, and Area Under the Curve (AUC). Results show that deep learning models, especially CNN and LSTM, did very well finding complex patterns about safety rules; hence, they did even better than traditional machine learning models. This study underlines the importance of using various models to improve prediction abilities about safety checks at construction sites, thus imparting useful information to safety managers and construction workers.