In order to ensure the reliability of the communication, the channel estimation technique is widely used in the OFDM system to reduce the bit error rate in the traditional broadband power line communication system. However, due to the need of importing a large number of pilot sequences, the channel estimation technology will take up valuable spectrum resources. Its implementation process is complex and will greatly reduce the effectiveness of communication. Therefore, we proposed a optimization algorithm of de mapping module for broadband PLC based on the Online Sequential-Extreme Learning Machine (OS-ELM). This algorithm is an online learning improvement algorithm of the traditional Extreme Learning Machine (ELM), which can combine batch processing and successive iteration to update training data and network parameters. This paper sets up a simulation model of the broadband power line communication system based on the actual data collected from the user electric meters in a residential district of Guangdong province. The simulation test is carried out under the measured 500m four-path channel and compared with the BP neural network and the traditional ELM. The experimental results show that the introduction of OS-ELM in a variety of SNR communication environments shows faster training speed and better anti-interference characteristics. Except for the extremely poor communication environment with low SNR, the algorithm can effectively improve the communication quality and reduce the bit error rate.