Data-driven load forecasting plays an important role in modern power systems. Due to the diversity of load characteristics in distribution transformer areas and the complexity of influencing factors, the prediction results of traditional methods lack interpretability. To address this issue, this paper proposes a deep learning interpretability method that considers the multidimensional characteristics of distribution network datasets. First, interpretability analysis is performed based on the prediction results of deep learning models. Then, similarity experiments and ablation studies are conducted from an internal perspective to examine the influence of dataset size and seasonality on the model’s forecasting performance. On this basis, a visualization and quantification method for load dispersion is proposed to verify, from an external perspective, the impact of dataset dispersion on model forecasting performance. Finally, by integrating dataset characteristics from both internal and external perspectives, a “dynamic matching” dataset optimization strategy is proposed. This strategy dynamically selects training samples from the database based on seasonal and dispersion features, thereby improving the representativeness of the samples. Test results show that the proposed method can identify key data that improve the prediction accuracy of the model, reduce the time cost of data collection, and enhance users’ trust in the forecasting results.