Short-term solar forecasting using deep long short-term memory recurrent network program | |
| Author | Tanawat Laopaiboon |
| Call Number | AIT Thesis no.ET-18-09 |
| Subject(s) | Solar energy--Forecasting Memory |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Energy |
| Publisher | Asian Institute of Technology |
| Abstract | Solar photovoltaic power generation is an intermittent renewable energy source. It is highly dependent on solar irradiance, cloud cover variability, temperature, atmospheric aerosol levels, and other atmosphere parameters. Accurate forecasting of solar power is crucial to short-term generation scheduling and on-line secure economic operation. This thesis proposes a short-term hourly solar forecasting technique using Deep Long-short term memory recurrent network (DEEP LSTM RNN)program considering time sequence data. Deep learning techniques are considered as one type of machine learning that can be used for load forecasting, solar forecasting and wind forecasting. The DEEP LSTM RNN have more advantages than shallow neural network in terms of model architecture and the training process. The network is constructed from combining Long-short term memory cell with recurrent network. The Long-short term memory, a kind of feedforward neural network with memory cell unit, is interconnected with an input layer and a hidden layer. With forget gate of the Long-short term memory, it can reduce time consuming and better process bad data during the training process. The recurrent network is a class of feedforward neural network, dividing the input feature into the time sequence. The architecture of recurrent network is designed with the interconnection of two hidden layers to better incorporate time sequence data than the shallow feedforward neural network. The training process of recurrent network is considered to be reinforcement training categories, combining with supervised learning and unsupervised learning. The unsupervised learning here is LSTM unit. The LSTM unit is used to pre-train the input data feeding to a hidden layer for reducing training time and avoiding vanishing gradient. By eliminating some redundant input, the model is considered to be supervised learning because we assign the target labeled data for testing comparison. The backpropagation process to the time of RNN is updating weights between interconnected two hidden layers to minimize the loss function (error between network output and desire output)suitable to process the time sequence than the shallow feedforward neural network. Therefore, combining LSTM with RNN is a deep learning model with complex interconnection of multi hidden layers. The input data used here include solar radiation from previous 7 intervals (time series data), day of the year, time of the day, temperature, and humidity. The simulation of hourly solar irradiation forecasting uses the solar irradiation, relevant meteorological and time series data as input which collected from previous 1 year (8760 hourly interval data)to forecast the sample data which is 7 days (91 hourly interval with removing night time hour)of the target current year. From back test simulation, the simulation results from DEEPLSTM RNN render a better performance than shallow neuron network in terms of root mean square error (RMSE), mean bias error (MBE), mean absolute percentage error (MAPE), mean absolute error (MAE) and correlation coefficient (COR).Comparing with Deep Belief network (DBN) and Auto Encoder Long-short term memory(AUTO-LSTM),our simulation results have lower RMSE, MAE and COR with slightly higher MBE than DBN and AUTO-LSTM. The proposed DEEP LSTM RNN) program is potentially viable for solar forecasting of utilities due to the higher accuracy. |
| Year | 2018 |
| Type | Thesis |
| School | School of Environment, Resources, and Development (SERD) |
| Department | Department of Energy and Climate Change (Former title: Department of Energy, Environment, and Climate Change (DEECC)) |
| Academic Program/FoS | Energy Technology (ET) |
| Chairperson(s) | Weerakorn Ongsakul |
| Examination Committee(s) | Singh, Jai Govind;Than Lin |
| Scholarship Donor(s) | The Bangchak Petroleum Public Company Limited, Thailand |
| Degree | Thesis (M.Eng.) - Asian Institute of Technology, 2018 |