<p>The dynamic and heterogeneous nature of Internet of Thing (IoT) edge environments, where billions of devices continuously generate diverse and time–varying workloads necessitate predictive mechanisms capable of accurately forecasting both task offloading requests and resource demands. While existing prediction–based offloading studies have primarily focused on forecasting task volumes or a single resource and in some cases both CPU and memory), none have jointly predicted offloading requests with task –level features alongside their associated resource requirements. To address this gap, we propose a hybrid multi task learning One-Dimensional Convolutional Neural Networks - Bidirectional Long-Short Term Memory (1D-CNN-BiLSTM)-attention model. The 1D-CNN component extracts local temporal patterns, the BiLSTM captures long–range dependencies, and the Attention mechanism emphasizes the most informative time steps and features. The model jointly predicts two categorical offloading parameters (task priority and delay tolerance) and four continuous resource–demand metrics (CPU request, memory request, maximum CPU usage, and maximum memory usage). A multi–objective learning strategy was employed, with classification targets representing task level semantics and regression targets estimating resource demands. Optimization was performed using a weighted combination of sparse categorical cross–entropy and mean squared error losses to effectively balance the heterogeneous objectives. The Google Cluster dataset was employed to train and evaluate the proposed model. Model performance was assessed using standard evaluation metrics, including MAE, MSE, RMSE, R², MAPE and ordinal-aware accuracy for categorical targets. The results indicate the superior performance of the proposed model consistently achieving the best results, with a minimum values of MAE =0.00014, MSE and RMSE values as low as 0.0001, and an R² score of 0.99, alongside a minimum MAPE of 2.30%. For classification tasks, the model attained the highest accuracy of 0.98. To ensure reliability, all experiments were repeated across five independent runs, and results are reported as mean ± standard deviation, demonstrating consistent performance with very low variance. Further, we have conducted a benchmarked based deep learning and classical machine learning models. Toward the end, when compared with prior state–of–the–art studies on workload prediction using the same dataset, the proposed model improved prediction accuracy by 6.4% in terms of R², thereby demonstrating its outperformance and advancement over existing approaches.</p>