Accurate forecasting of blood glucose levels (BGL) in individuals with type 1 diabetes mellitus (T1DM) is crucial for effective disease management. This study focuses on advancing BGL prediction by performing a comparative evaluation of single machine learning models and heterogeneous ensembles. The main objective of this study is to determine the optimal model configuration for accurate and reliable blood glucose predictions by examining various optimization strategies. For this purpose, several machine learning techniques were used, including Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), Support Vector Regression (SVR) and Deep Belief Networks (DBN). All these models were incorporated into various ensembles and fine tuned using Particle Swarm Optimization (PSO), Random Search (RS) and Bayesian Optimization (BO). We evaluated these models using metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Magnitude Relative Error (MMRE) and Predictive Level at 25% (PRED). Our findings demonstrated that PSO fine-tuned models, particularly those using GRU, performed exceptionally well. Most significantly, the PGRU model achieved a RMSE of 7.87 mg/dl. Ensembles optimized using RS and PSO, in particularly ERME and EPAV, outperformed all others, with RMSE values of 12.25 mg/dl and 12.26 mg/dl, respectively. Statistical validation through the Scott-Knott test (SK) and Borda Count (BC) analysis corroborated that PSO significantly enhanced accuracy in both individual and ensemble models. In conclusion, this study underscores the high efficacy of PSO in optimizing BGL forecasting models. Additionally, heterogeneous ensembles offer more consistent and reliable predictions than single models, thereby providing a promising solution for real-time diabetes management.