With the development of the carbon trading system, accurate forecasting of power consumption is crucial for energy management. For power load forecasting without key information such as weather information, phase space reconstruction technique of chaos theory is adopted first to deal with historical load time series. Permutation entropy is used to validate the chaotic characteristic. 8 kinds of machine learning models are used to forecast and make comparisons, These models are: 4 kinds of neural network, 3 kinds of statistical machine learning and 1 kind of benchmark. Secondly, two optimal models, extreme learning machine (ELM) and extreme gradient boosting (XGBoost), are combined by a grey relational degree method to construct an ELM-XGBoost model. Finally, ELM-XGBoost is applied to forecast with different time scales from one day to one week. Results show that forecasting accuracy decreases with the increase of forecasting time scale. In daily load forecasting, the accuracy of ELM-XGBoost model is improved with a better application effect. This work is supported by the National Key Research and Development Program of China (No. 2020YFB1506802).