Accurate short-term power load forecasting is critical for the stable and efficient operation of smart grids. However, due to the influence of various external factors, load data often exhibit strong non-stationarity and random fluctuations, making high-precision forecasting a challenging task. To address this problem, this paper proposes a short-term power load forecasting model based on an Extreme Learning Machine (ELM) optimized by a multi-strategy Improved Coati Optimization Algorithm (ICOA). The proposed improvements include: Tent mapping initialization to enhance population diversity and convergence speed; a random opposition-based learning mechanism to promote balanced population distribution and further accelerate convergence; and the incorporation of Levy flight and sine-cosine strategies to enhance global exploration and local exploitation, thereby improving the algorithm’s robustness against local optima. Ablation experiments and Wilcoxon rank-sum tests validate the effectiveness of each strategy, showing that the improved ICOA significantly outperforms the original COA and several other commonly used algorithms. Finally, the optimized ICOA is integrated with ELM to construct a hybrid forecasting model. Experimental evaluations on actual 2024 power load data from a province in China demonstrate that the proposed ICOA-ELM model achieves superior forecasting accuracy and faster convergence, effectively reducing prediction error and enhancing the reliability of short-term power load forecasting.