The rapid progress in machine learning (ML) has opened countless possibilities across various fields, yet predicting average temperature (AT) accurately remains a significant hurdle in climate modelling. This study introduces a novel approach called the Hybrid Asymmetric Huber Loss Function-Based Extreme Learning Machine (AHELM) to tackle this challenge, specifically focusing on the Kokernag station in the Jhelum River Basin, India. The goal is to improve temperature prediction by combining the strengths of Extreme Learning Machine (ELM) models with a custom Asymmetric Huber Loss Function, which helps the model handle outliers more effectively and boosts its overall accuracy. To train the model, historical temperature data was used, with different time-lagged values (such as ATt-1, ATt-2, ATt-3, and ATt-4) serving as input features. This approach ensures the model captures temporal patterns in temperature changes, optimizing its performance. The proposed AHELM model underwent evaluating predictive accuracy and reliability testing by statistical comparison with traditional ELM using R2, RMSE and NSE metrics. The results were promising: the AHELM model outperformed the standalone ELM, achieving an R2 of 0.9812 and an NSE of 0.9756. The hybrid model produces superior accuracy in temperature forecasting together with enhanced forecasting reliability according to the computed metrics.