Development of deep learning predictive models for short-term solar radiation forecasting: case study in Vietnam

Development of deep learning predictive models for short-term solar radiation forecasting: case study in Vietnam

Vietnam is a developing country with a projected high economic growth. The energy sector plays an essential role in its socio-economic development, with an average of 5% increase in annual energy demands. As a result of its growing energy consumption, the nation is hugely reliant on imported fossil fuels. In the foreseeable future, oil-based fuels are expected to become somewhat limited and even too exhausted. Coal resources, in general, are not easy to exploit or utilise due to notable climate change, economic and technical limitations. This situation increases pressure on national energy security in Vietnam. Considering the high solar radiation footprint in Vietnam, future assessments of the availability of renewable energy resources, through research and development initiatives and modelling studies, such as the one performed in this research study, can effectively provide meaningful information to Vietnamese policymakers in seeking alternative energy resources to replace the finite oil and coal reserves. Freely available renewable resources, such as solar energy, can be ideal sustainable solutions to satisfy the needs of energy security into the future. Research into solar energy forecasting has great potential in the search for alternative energy resources.

Since the availability of solar radiation at the earth’s surface is directly proportional to the harvested solar power at any specified location, the forecasting of global solar radiation is essential for continuous monitoring and supply management of solar energy through solar generation systems. Therefore, the forecasting of solar radiation has been studied extensively in literature, using many forecasting methods, but these methods have been largely based on physical and statistically driven models. In spite of their usage, these methods come with certain limitations, such as the underlying assumptions of initial conditions, mainly with physics-based models. The concept of persistence to forecast future solar radiation, as used in previous methods, can also be a challenging factor to attain good accuracy. In contrast, physical models may not adequately address issues of stochasticity (i.e. rapid change in solar radiation or its predictor variable, e.g. cloud cover) datasets, or the issues of data non-stationarity. As an alternative forecasting method, machine learning approaches, which capture historical behaviour of solar radiation or its predictor variables to model utilising artificial intelligence algorithms are now becoming prominent. Such a model is constructed and trained to learn the patterns in historical solar radiation (and related) datasets, to build a non-linear mapping scheme between the antecedent (i.e. lagged) inputs and the target (i.e. solar radiation) dataset. This thesis focuses on developing and evaluating latest machine learning models for global solar radiation forecasting in the context of Vietnam, where the potential utility of machine learning models has not yet been fully explored.

The primary aim of this Master of Science (MSCR) thesis is to develop new and scientifically verified models to resolve the challenges in solar radiation forecasting, addressing the issues of complexity in predictor and the target datasets while also capturing the non-linear behaviour of solar radiation. To pursue this aim, the study is based on Long Short-Term Memory (LSTM) network algorithm built for global solar radiation forecasting. In some other studies, LSTM has attained superior capability in learning the long- and short-term dependencies in historical datasets, by employing memory cells that determine the importance and distinguish between the important and unimportant data features through their input, forget and output network gates. Consequently, the essential data features are then used to build the feature engineering and data pattern learning process in the resulting LSTM model. This research, therefore, utilises the LSTM model to forecast the hourly solar radiation and also attempt to capture the dependence between consecutive hours on the same day, as well as the long-term (e.g. seasonal) behaviours to be learned efficiently.

This MSCR thesis, presented as a synthesis of two journal publications, has adopted an LSTM model incorporating a data pre-processing technique based on the Robust Local Mean Decomposition (RLMD). The research study area has focussed on solar energy belts, which are the critical zones in Vietnam for major solar energy projects. The aims of this research thesis are as follows: (1) To develop a near real-time solar radiation forecasting model using the LSTM algorithm applied to multiple time-step forecast horizons (this work has been reported in Journal Paper 1); (2) To further improve the method in the first aim and build a hybrid forecasting model utilising a data pre-processing technique based on the RLMD method, and further evaluate the forecasting performance of the resulting hybrid RLMD-LSTM model using half-hourly global solar radiation forecasting (this work has been reported in Journal Paper 2 – under review). The overall results of this study show the LSTM model can be used as a useful utility in global solar radiation forecasting at near real-time horizons. The findings of this study can have important implications for renewable energy feasibility studies, and also help in several areas where data-driven decisions may require some of the best practice forecasting techniques.

In synopsis, the predictive models developed in this MSCR thesis will provide significant benefits to solar energy generators, authorities for energy operation and distribution, through new and improved solar radiation forecasting tools. Energy forecasters can therefore adopt these novel methods, to address the issues of non-linearity and the non-stationarity in energy usage, by constructing real-time forecasting tailored for energy industries and other stakeholders.