Electronic Thesis and Dissertation Repository

Thesis Format

Integrated Article

Degree

Master of Engineering Science

Program

Electrical and Computer Engineering

Collaborative Specialization

Artificial Intelligence

Supervisor

Grolinger, Katarina

Abstract

To combat negative environmental conditions, reduce operating costs, and identify energy savings opportunities, it is essential to efficiently manage energy consumption. Internet of Things (IoT) devices, including widely-used smart meters, have created possibilities for sensor based energy forecasting. Machine learning algorithms commonly used for energy forecasting, such as FeedForward Neural Networks, are not well-suited for interpreting the time dimensionality of a signal. Consequently, this thesis applies Sequence-to-Sequence (S2S) Recurrent Neural Networks (RNNs) with attention for electrical load forecasting. The S2S and S2S attention architectures commonly used for neural machine translation are adapted for energy forecasting. An RNN enables capturing time dependencies present in the load data, while the S2S RNN model strengthens consecutive sequence prediction by combining two RNNs: encoder and decoder. Adding the attention mechanism to these S2S RNNs alleviates the burden of connecting the encoder and decoder. Presented experiments compare a regular S2S model and four S2S attention models with two baseline models, the conventional Non-S2S RNN and a Deep Neural Network (DNN). Furthermore, each RNN model was evaluated with three different RNN-cells: Vanilla RNN, Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) cell. All models were trained and tested on one building-level electrical load dataset, with five-minute incremental data. Results showed that the S2S Bahdanau et al. attention model was the dominant model as it outperformed all other models for nearly all forecasting lengths.

Summary for Lay Audience

Energy consumption has been continuously increasing due to the rapid expansion of high-density cities, and growth in the industrial and commercial sectors. To combat negative environmental conditions, reduce operating costs, and identify energy savings opportunities, it is essential to efficiently manage energy consumption. Internet of Things (IoT) devices, such as widely used smart meters, are capable of measuring and communicating data about energy use; thus, they have created opportunities for improved energy management as well as for energy forecasting. Machine learning techniques build a mathematical model based on this historical data in order to make predictions or perform a different task. The common machine learning algorithms used for energy forecasting are not well-suited for time series problems. Consequently, this thesis applies Sequence-to-Sequence (S2S) Recurrent Neural Networks (RNNs) with attention for electrical load forecasting. Specifically, S2S and S2S attention models from neural machine translation are adapted for energy forecasting. RNNs enable capturing time dependencies present in the consumption data, while the S2S RNN strengthens consecutive predictions by combining two RNNs: encoder and decoder. The first RNN (encoder) reads the data, passes information to the second RNN (decoder), and the second one predicts the energy consumption. Further, an attention mechanism was added to the overall model, which helps connect encoder and decoder RNN and allows the decoder RNN to pay attention to specific parts of the encoder RNN. By using these techniques, the proposed approach can improve energy consumption forecasts. Presented experiments compare seven models: a regular S2S model, four S2S attention models, and two baseline models, the conventional RNN and a Deep Neural Network (DNN). All models were trained and tested on one building-level electrical load dataset, with five-minute incremental data. Results showed that the S2S Bahdanau et al. attention model was the dominant model as it outperformed all other models for nearly all forecasting lengths.

Share

COinS