The growing IoT network of linked devices creates massive volumes of data that can be analyzed and utilized to make choices. Machine learning algorithms, which need plenty of data to learn, struggle with IoT data's unpredictability and dispersion. Using federated learning, a novel machine learning method, many devices may develop a global model without exchanging raw data with a central server. We introduce a stochastic gradient descent (SGD)-based federated learning method for scalable IoT analytics in this paper. Our method uses smart meters and other IoT devices to develop a worldwide energy demand model. We advocate employing a distributed SGD approach to train smaller components of the global model on several devices at once. We used smart meter readings to show that our technique is more exact and scalable than centralized learning methods. Because the raw data is saved locally on the devices rather being shared with a server, our solution protects privacy. Our proposed approach to IoT analytics difficulties uses federated learning to solve them in a distributed and private way.