Synthetic Temperature Prediction Using an Intuitionistic Fuzzy Neural Network by Particle Swarm Optimization

Accurate temperature prediction is essential for the assessment of intelligent data-driven prediction methods under nonlinear and uncertain operating conditions. However, uncertainties in the environmental variables and nonlinear relationships between the input variables pose challenges to conventional prediction methods. This study is intended to show that the IFNN–PSO can be used to demonstrate proof of concept in the application of synthetic temperature prediction. This paper proposes a novel system, which is the combination of intuitionistic fuzzy representation, neural network learning, and PSO-based parameter optimization, aimed at modeling a complex relationship between the temperature-related synthetic variables. The model takes outside temperature, inside temperature, humidity, time of day and the number of people as input parameters and predicts the desired indoor temperature when these parameters are set to a synthetic set. The IFNN has 5 input nodes, 15 hidden units and 2 output nodes for membership and non-membership degree which give the total trainable parameters of 122. These parameters are optimized using PSO by minimizing the mean squared error. The dataset was split into 70% as a model development set, and 30% as a protected test set; five-fold cross-validation was used only for model selection and hyperparameter tuning within the model development set, and the protected test set was used only towards the end to assess the final performance. On the protected test set, IFNN–PSO achieves an MSE of 0.0853, RMSE of 0.2920, MAE of 0.2314, and (R2) of 0.8784, outperforming the ANN baseline, which obtains MSE = 0.0968, RMSE = 0.3111, MAE = 0.2446, and (R2) = 0.8621. The model produced by IFNN–PSO has the minimum error values among the tested models and a competitive goodness of fit in comparison to the benchmark models. These findings support IFNN–PSO as a competitive method for synthetic temperature prediction and provide a basis for future validation of its potential contribution to energy efficiency in buildings.