As renewable energy sources become more prevalent, due to their variable and interment nature, having a timely and accurate forecast of production will be ever more vital for their effective use. Inaccurately forecasting output can cause the generator to deviate from their commitment to the grid. In order to maintain stability the grid operator is forced to intervene, taking action to balance the grid. The cost of this action, the balancing cost, is passed onto the offending producer. Since the grid, in the short run, usually trades in 30 minute blocks having a prediction only a few hours into the future can be enough to make an impact. In the case of solar, generation is variable and peaks during the middle of the day. The main source variability in power output is the amount of solar irradiance that falls on the generator. Some of the biggest factors that effect irradiance are atmospheric / meteorological. Having an accurate picture of the weather, cloud location, density, direction and speed of movement would help facilitate the creation of a predictive solar irradiance model. Today commercial solar farms are continuously producing vast amounts of data. Accurate measures of solar irradiance, power output as well as detailed telemetry for each component. This can amount to 100,000s of data points per plant being generated every day. The volume and velocity of data can necessitate the use of big data tools such as Apache Spark. As previously stated, a major factor in solar production is the weather. Commercial providers claim to offer accurate observations for virtually any point on the planet. One can also access rich radar and satellite imagery. This type of data lends itself well to machine learning (ML). There are many interesting ML approaches that have been show to work leveraging the sequential properties of this kind of data to make predictions. There are many examples in the literature of people applying ML techniques to predict solar irradiance. Alzahrani et al have demonstrated use of deep learning, in the form of a recurrent neural network, to forecast solar irradiance [1]. Marquez et al have used meteorological data and an ANN to predict solar irradiance up to 6 days out [2]. Various optimisation techniques also exist within the literature for a broad range of domains. Xue et al describe a simulation-based optimisation for perishable food, minimising waist whist ensuring enough supply to meet customer demand [3].
Aims and Objectives The main aim of the research is to develop novel big data learning and optimisation techniques. Specifically for the practical application of predicting, power output for arbitrary sized solar arrays (either domestic or commercial) and optimising power use from the array (storage / selling to the grid). Specifically this will involve: Power Prediction: Design and develop a model to predict solar irradiance. Design and develop a system to take irradiance predictions and convert to power output for a given solar array. Power Optimisation: Design and develop a system to model / predict power use and prices throughout the day. Design and develop a system to optimise how power from the solar array is used conforming to various constrains (e.g battery charge cycles, gird buy / sell prices, predicted future power output).
References [1] Ahmad Alzahrani, Pourya Shamsi, Cihan Dagli, and Mehdi Ferdowsi. Solar irradiance forecasting using deep neural networks. Procedia Computer Science, 114:304 - 313, 2017. [2] Ricardo Marquez and Carlos F.M. Coimbra. Forecasting of global and direct solar irradiance using stochastic learning methods, ground experiments and the NWS database. Solar Energy, 85(5):746 - 756, 2011. [3] Ning Xue, Dario Landa-Silva, Grazziela Figueredo, and Isaac Triguero. A simulation-based optimisation approach for inventory management of highly perishable food. 02 2019.