Improving boiler thermal efficiency plays a very important role in the economic development of power plants. In order to implement a real-time improvement in the boiler thermal efficiency, a precise and rapid online model of the thermal efficiency is required. The present paper presents an effective machine learning method called the Online Least Square Fast Learning Network (OLSFLN) to build a prediction model for 300 MW coal-fired boiler thermal efficiency. Experimental results demonstrate that the proposed OLSFLN could predict the boiler thermal efficiency with high accuracy and outperform in learning ability, generalization ability and repeatability under various boiler operating conditions than other state-of-the-art algorithms.