As the proportion of distributed renewable energy in power grids continues to rise, load fluctuations are becoming more complicated. Meanwhile, factors such as electricity prices and human activities have been further increasing the complexity of load fluctuations, making load forecasting more challenging than ever before. To address this challenge, this paper proposes a load forecasting method based on Convolutional Neural Networks-Long Short-Term Memory-Attention (CNN-LSTM-Attention) network, using a variety of multi-source data including historical load, temperature, electricity price, solar and wind power generation, radiation intensity and the types of days. First, the grayscale analysis method is applied to analyze the correlation between load and other data, determining the input features of the load forecasting model. Then, the input sequence of the model is created from three time dimensions including adjacent time slots, short cycles, and long cycles. Finally, a CNN-LSTM-Attention network is built for training the model. Experiments conducted on real energy data from the Austrian region, demonstrate that the proposed model achieves higher forecasting accuracy compared to other models. As a result, the model proposed in this paper can effectively utilize time series information across multiple time dimensions, and adapt to the complexity of load fluctuations by employing various type of input feature, significantly improving the accuracy of load forecasting.