emmxl / -machine-learning-to-predict-energy-consumption

This a machine learning model to predict electricity consumption, the dataset used was from Finland.

Energy Consumption Prediction using Machine Learning

Time Series Forecasting (LSTM)

This research was done as part of a research seminar course at LAB University of Applied Science (fall 2022). The report was written using the thesis template of LAB University.

Abstract

The purpose of this research was to predict energy consumption using the data of Finland's transmission system operator. The objective of this project was to test if a machine learning model can yield good enough results in a complex forecasting problem, exploring machine learning techniques and developing a data-driven model for forecasting energy. The data contained 6-year hourly electrical consumption in Finland, and it is a univariate time series, as it is seasonal. We used a long-short term memory (LSTM) model to train the data. The model was evaluated using root mean squared error (RMSE) to be directly comparable to energy readings in the data. The result shows that electricity consumption can be predicted using machine learning algorithms so we can use the results to deploy renewable energy, plan for high/low load days, and reduce wastage from polluting on reserve standby generation.

Model Implementation

The data was imported from Finland's transmission system operator as a CSV file and then exported to a GitHub repository. There was a total number of 52965 observations and 5 variables in this dataset and no missing values were found. The minimum load volume is 5341 MWh, and the maximum load volume is 15105 MWh along with an average volume of 9488.750519 MWh. The data is univariate time series, where there is a need for one column to present time and another one to present energy consumption. For predicting day consumption, data were down-sampled using resample function. This function changed the data from hourly frequency to daily frequency. Training the LSTM model was done using the training set and the validation dataset for testing the results through the training process. The learning algorithm worked through the entire training dataset 60 times (Epoch), and the model weights were updated after each batch where the batch size is 20.