The load data of distribution transformer area not only exhibits autocorrelation as time-series data but also shows non-stationarity due to the influence of environmental factors. Therefore, the prediction accuracy is not only related to the structure of the prediction model but also related to the time series characteristics of the input data. In order to improve the prediction accuracy of the model, this paper proposes a multi-granularity load forecasting model based on chaotic time series analysis and kernel extreme learning machine. To deal with the non-stationary characteristics of the load data, the non-stationary original signal is converted into a series of relatively stable sub-signals through the variational modal decomposition algorithm. Regarding the autocorrelation characteristics in the load data, the chaotic time series analysis method is used to solve the size of the time window of each modal when the modal is input into the prediction model. By constructing a forecasting model based on multi-granular kernel extreme learning machine, the adverse effects of non-stationarity and autocorrelation in the load data on the load prediction are reduced, thus improving the prediction accuracy of the model. The results show that the load forecast accuracy is affected by the size of the time window of the input data, and the optimal time window for different modal components is different. The chaotic phase time series analysis method can estimate the optimal time window size of each modal component, effectively improving the prediction accuracy of the kernel extreme learning machine.