The accurate prediction of power load is a prerequisite for maintaining the supply-demand stability of the power system. To improve the accuracy of load forecasting, we proposed a novel fluctuation- and smooth-channel-based deep learning model for load forecasting. First, the multidimensional information-aware sensor technology is used to collect wind speed, temperature, and humidity information through the analog-to-digital converter (ADC) digital-information-acquisition, filter-circuit-noise-elimination, and amplification circuits. The microcontroller in the sensor is employed to process the output of the characteristics of the impact of the load data information. Then, the interquartile range was employed to detect the abnormal values in the load data, and the missing values caused by their removal were filled and eliminated by cubic spline interpolation to enhance the quality of the load data. Second, the long short-term memory (LSTM) model based on a fluctuation channel and a smooth channel was constructed, which can autonomously distinguish the fluctuation period from the smooth period in the data, fully exploiting the fluctuation information in the fluctuation period and the subtle changes in the smooth period. Additionally, an improved catch fish optimization algorithm was specifically designed to optimize the hyperparameters of the prediction model, enhancing its ability to characterize complex load sequences. Finally, the proposed method and model were validated through case studies. The results demonstrated that compared with existing load prediction models, the proposed model achieved a mean absolute percentage error below 3% and a goodness-of-fit exceeding 98%, effectively capturing the fluctuation trend of complex load sequences.