(An) accelerated streaming data processing scheme based on CNN-LSTM hybrid model in energy service platform

Abstract

Energy platforms help reduce energy costs, provide reliable power, and maintain customers by measuring, analyzing, and controlling composite systems. It can also have several energy-related service systems, such as planning for provisioning power resources, pricing and billing of energy, and managing energy storage. Control of the energy system is difficult because variables such as price and energy consumption exhibit seasonality and are not consistent. Therefore, making accurate predictions about future energy demand and supply may enhance the stable control of this energy service system. While traditional time-series forecasting methods have not been accurate, deep learning has shown high performance by extracting meaningful information and hidden patterns of incoming data from distributed and heterogeneous sensing generating massive amounts of data streams. However, the recent increase in the role of renewable energy production, from sources such as wind power and solar power, being dramatically variable or intermittent in supply, hinders the direct application of the traditional deep learning. Many deep learning applications in real-world involve processing of data streams that change over time and theoretically-infinite set of data, and as a result, learning models should be continuously updated to reflect the recent trend and distribution in data. In this thesis, we used Convolution Neural Network Long Short-term Memory Network (CNN-LSTM), using Convolutional Neural Network (CNN) layers for feature extraction on input data combined with LSTMs to support sequence prediction, to demonstrate the excellence of deep learning in forecasts of renewable data and electricity consumption data. Additionally, we proposed the deep learning-base system which consists of two main phases: re-training and deployment procedure. The former is to determine the amounts of data streams and the number of iterations in the process of updating parameters to reflect newly generated data at the right time and keep the model up-to-date with new training data, thereby enhancing accuracy by incorporating data into models used for inference. The latter is to check whether the trained model is appropriate for serving using Euclidean Distance. By not changing the model used for inference if the Euclidean distance between the currently used model and the newly updated model is smaller than the threshold, the time including the time used for memory loads can be saved. This model update approach has shown that it reduces the amount of computing and time needed for learning rather than online learning, in which data is incorporated instance by instance, leading to lower learning costs.