In the next few years the persistent growth of photovoltaic penetration will be the most important change in the Hungarian energy sector. For this reason, the aim of this work is to investigate the daily load prediction difficulties from the TSO perspective. Firstly, I reviewed the relevant trends of the photovoltaic industry both globally and locally in the last 10 years. Then I focused on the challenges of scheduling of the conventional plants caused by the volatile energy production of the solar plants. I made a short-term prediction to the expected growth of the solar penetration using previous trends in Hungary.
In the second part of my work I simulated the effect of the growing photovoltaic energy production to the daily load curve. For this reason, I had to create a model which is able to generate load prediction with an 5% accuracy. My neural network was trained by the daily load values and environmental parameters (e.g. temperature) in the period of 2010-2016, when the solar penetration was negligible. All in all, the network was able to generate day ahead prediction without the distortion effects of the photovoltaic generation. Then it was used to make prediction to the year of 2017 and I investigated the tendencies of the prediction error.
It was needed to acquire a new scientific field for me to take over the tasks of this work, because I have never dealt with Machine Learning and programming tasks. So, this paper introduces the developing process both the used neural models and my personal skills. Firstly, I have used Matlab built-in Neural Fitting Toolbox for training, but the results were not accurate enough. Due to the closed source software and the few parameters I had to implement the model in Python, and I was able to refine the resolution of the prediction. Unfortunately, it was not enough for stable consequences.
Finally, I have used Long Short-term Memory, which is used primarily for predictions, and I was able to define the effect of solar energy production.