Abstract: Given the deep and intractable nonlinear coupling among multi-source heterogeneous data, coupled with the nearly chaotic random fluctuations at both the source and load ends, traditional data-driven models are highly susceptible to the dilemma of prediction "stalling" during "cold starts" or data interruption scenarios. To address this issue, this study proposes a novel paradigm for microgrid load situational awareness that integrates mechanism-augmented data generation with deep spatiotemporal interaction. Specifically, a mechanism simulation model with multi-dimensional mapping relationships is designed to inversely generate high-fidelity physical scenarios for data augmentation, followed by the introduction of an adaptive Hampel filter as a data "purifier". Building upon this data foundation, a targeted CNN-BiLSTM cascaded architecture tailored to physical coupling characteristics is employed. A parameter-customized 1D-CNN is utilized to keenly capture local short-term sudden changes among multiple variables, working in synergy with a BiLSTM to deeply mine the bidirectional long-range dependencies within the time series. Experimental results demonstrate that the proposed method achieves a Mean Absolute Percentage Error (MAPE) as low as 4.90% and a coefficient of determination (R2) up to 0.9102 on the test set, reducing the prediction error by 20% compared to the traditional LSTM model. This effectively resolves the core pain points of traditional models, such as poor adaptability to small samples and prediction stalling, thereby providing reliable technical support for the optimal scheduling and safe, stable operation of microgrids.