Project overview
The rapid establishment of the Internet of Things (IoT) provides a unique situation to reach a massive amount of data to create valuable insights for the businesses. Efficient analysis of communicated sensor data across IoT networks helps to enhance the operations of the supplier businesses such as load balancing, supply planning, and especially marketing strategies. It is also used for conservation of energy and balancing the energy towards green sources. Therefore, the popularity of time series forecasting for sensor data is growing with the number of utilities moving to smart systems. However, high-frequency data collection and a vast volume of data make sensor data prediction more challenging which requires more extended processing and training times. Thus, machine learning methods in the IoT system need to be efficient in handling large amounts of dynamic data, which can be difficult for traditional approaches.
One of the crucial aims of implementing data analytics in IoT is comprehending data patterns in sensor data. In addition to forecasting performance and relevant demanding load, anomaly detection, which is referred to as the discovery of unexpected events, assists in discovering and reforming unplanned situations swiftly. Consequently, training a model to predict the behaviour of sensor data and identifying anomalies will enhance sensor maintenance leading to operational efficiency.
On the other hand, by the continuous increase in computational power and integration of machine learning methods with the forecasting problems, there are several approaches to cope with big-time series data with shorter response times and possibly higher performances and robustness of the models. One of the techniques in machine learning is transfer learning that extracts the information from previously learned domains and implements it for different but related target tasks. This approach with reducing training computation enhances learning model performance. Consequently, it has the potential to offer an appropriate solution for time series forecasting problems that have to deal with the volumes and training times required. The fundamental idea behind transfer learning is that instead of separate training models for each time series from scratch, one model can be trained on multiple time series that are related and similar.
The general application framework of transfer learning on time series forecasting is missing, as well as the special focus on sensor data that is a fertile application area for it. From this point of view, we will conduct experiments on the integration of transfer learning with time series forecasting models for sensor data obtained through smart meters to investigate the improvements in models’ performance (in terms of accuracy and speed). In this research with implementing a recurrent neural network (RNN) in a transfer learning framework, we are looking for to find out the answer of following questions:
1. Does transfer learning enhance time series forecasting in sensor data analysing?
2. How much would RNN with the proposed transfer learning framework improve detecting abnormal events in sensor data?
3. Is it possible to create certain groups/clusters of time series that are optimal for transfer learning?
One of the crucial aims of implementing data analytics in IoT is comprehending data patterns in sensor data. In addition to forecasting performance and relevant demanding load, anomaly detection, which is referred to as the discovery of unexpected events, assists in discovering and reforming unplanned situations swiftly. Consequently, training a model to predict the behaviour of sensor data and identifying anomalies will enhance sensor maintenance leading to operational efficiency.
On the other hand, by the continuous increase in computational power and integration of machine learning methods with the forecasting problems, there are several approaches to cope with big-time series data with shorter response times and possibly higher performances and robustness of the models. One of the techniques in machine learning is transfer learning that extracts the information from previously learned domains and implements it for different but related target tasks. This approach with reducing training computation enhances learning model performance. Consequently, it has the potential to offer an appropriate solution for time series forecasting problems that have to deal with the volumes and training times required. The fundamental idea behind transfer learning is that instead of separate training models for each time series from scratch, one model can be trained on multiple time series that are related and similar.
The general application framework of transfer learning on time series forecasting is missing, as well as the special focus on sensor data that is a fertile application area for it. From this point of view, we will conduct experiments on the integration of transfer learning with time series forecasting models for sensor data obtained through smart meters to investigate the improvements in models’ performance (in terms of accuracy and speed). In this research with implementing a recurrent neural network (RNN) in a transfer learning framework, we are looking for to find out the answer of following questions:
1. Does transfer learning enhance time series forecasting in sensor data analysing?
2. How much would RNN with the proposed transfer learning framework improve detecting abnormal events in sensor data?
3. Is it possible to create certain groups/clusters of time series that are optimal for transfer learning?
Staff
Lead researchers
