Smart cities commonly use machine learning to forecast the electricity load to help the smart grids to allocate the power in advance, as high forecasting accuracy is necessary to ensure grid stability and efficient resource allocation. However, popular methods, especially models based on artificial neural networks, must rely on imputation as they cannot directly process missing values. Although we can use masking or padding as the most cost-effective alternatives, the effectiveness decrease as the missing value increase. Moreover, due to the centralized design, most existing methods suffer from a "single point of failure" vulnerability. With the fast adoption rate of digital technology within urban environments, smart cities are becoming more profitable targets for hackers, which cause large-scale services disruption via denial-of-service attacks causing packet loss or bringing the server offline. To solve this, we proposed a distributed forecasting network resilient against downtime and missing values. Compared to the previous best machine learning model, which got coefficient of determination scores of 0.9442 when 50% of the data is missing, our proposed method achieves 0.9451 in regular operation, which changes to 0.9453 when switching to the distributed forecasting method.