[Objective] Constrained by practical operating conditions， load forecasting often faces the dual challenges of low data quality and variable load distribution. To address this， an ultra-short-term load forecasting method considering historical data missing and load temporal heterogeneity is proposed in this paper.[Methods] First， to reconstruct missing time-series data such as load， a Convolutional Neural Network-Bidirectional Long Short-Term Memory （CNN-BiLSTM） neural network is embedded within the Generative Adversarial Imputation Network （GAIN） to capture the spatiotemporal dependencies. During training， to prevent the model from focusing solely on the imputation error of observable values， random noise is introduced to replace partial observable values， enabling explicit measurement of the generation bias corresponding to the aforementioned noise. Second， to reveal load heterogeneity at finer temporal scales， clustering is applied to load input samples for the forecasting model in the training set to identify different load distribution patterns. Finally， a unified forecasting model is fine-tuned according to different pattern-specific samples， constructing personalized sub-models. During online forecasting， the most suitable sub-model is dynamically selected based on the similarity between the current load input sample and the centers of each pattern， eliminating reliance on external information such as calendar and weather data.[Results] Case studies demonstrate that the proposed missing data reconstruction method achieves lower reconstruction errors compared with traditional methods. Based on this， the constructed forecasting model yields higher prediction accuracy. Incorporating the construction and selection strategy for personalized forecasting sub-models further decreases forecasting errors.[Conclusions] Experimental results confirm the practical engineering value of the proposed method in real-world operational scenarios.