The utilization of renewable sources of energy is growing all over the world due to pressure for sustainable solutions. It brings benefits to the environment, but also adds complexity to the electricity grid, which faces energy balancing challenges caused by an intermittent production from this kind of generation. Having a good energy prediction is essential to avoid losses and improve the quality and efficiency of the energy systems. There are many machine learning (ML) methods that can be used in these predictions; however, every consumer is different and will behave in a distinct way. Therefore, the objective of this article is to compare the application of different ML methods, aiming to predict PV energy production and energy consumption for residential users. Four different ML methods were applied in a real dataset from the RESPOND project: Linear Regression, Decision Forest regression, Boosted Decision Tree Regression and Neural Network. After the simulation, the predicted values were compared against the real data, considering 150 days of measurement from two Irish houses. Overall, all the algorithms applied achieved mean errors below 14%, but the Boosted Decision Tree overperformed, with mean errors of 2.68% and 10% for energy consumption and energy production prediction, respectively.
Energy generation through renewable sources is becoming popular in recent years. The
European Union (EU) targets to achieve about 20% of renewable energy production in
2020 [
Understanding and predicting how electricity network works, including distributed
generation from renewables, is essential in this new framework, as it can bring benefits
to the utilities. Contemporary solutions for energy balance, such as backup fossil
powerplants [
To support this new trend, some data-driven methods for energy production and
consumption prediction have been arising, ranging from statistical models to complex
Machine Learning (ML) algorithms. These aim to find correlations and meaning among
variables in large datasets. Although more than 80% of the previous studies about
energy consumption prediction have been carried on non-residential customers [
In summary, the main contribution of this work is to assess Machine Learning techniques for energy prediction and to deploy a simulation environment, aiming to provide the following predictions for a hypothetical EMS: 1. Photovoltaic (PV) energy generation. 2. Residential energy consumption for a small group of houses.
The rest of this paper is organized as follows: Related work, which shows the related work regarding energy production and consumption prediction. Machine Learning section provides information about the principles and techniques applied in this research. Environmental Setup describes how the environment monitoring has been structured and explains the dataset preparation stage. Results section presents all the relevant outputs of our experiments, comparing the prediction methods and real data. Finally, Conclusions and Future works recaps the main points of the paper, introducing ideas for future work. 2
Electricity is a development indicator, it boosts country’s economy and brings comfort
to our homes, improving quality of life in most of daily tasks. However, it is also
strongly associated with CO2 emissions, where buildings represent 36% of the total
produced gas in the EU [
Amasyali and El-Gohary (2018) carried out an extensive review [
In a different approach, Naji at al. [
Regarding energy production prediction, Das at al. [
The authors in [
Machine Learning is a subfield of computer science that is classified as an artificial intelligence method. It can be used in several domains and one of the advantages is the capability of solving problems which are impossible to be represented by explicit algorithms. Some of the ML methods are regression based, which can be widely used to create projections about future, with the objective to predict a numeric target.
The best ML model will rely on the equilibrium between predicted error and complexity of the system. Depending on the database particularities, a complex model may result in a greater error than using a simple model, as shown in Fig. 1.
All methods to be presented here are regression-based. The following subsections describe the methods applied to our proposed prediction model. 3.1
Linear regression is a statistical method, which has been adopted for using in ML. Spite
of being one of the simplest models for a basic predictive task, this method also tends
to work well on high-dimensional sparse datasets [
Decision trees are non-parametric models that perform a sequence of simple tests for each instance, traversing a binary tree data structure until a leaf node (decision) is reached. The advantage of decision trees is that this method is efficient in both computation and memory usage during training and prediction.
Decision Forest model consists of an ensemble of decision trees. Each tree in a
regression decision forest outputs a Gaussian distribution as a prediction. An
aggregation is performed over the ensemble of trees to find a Gaussian distribution
closest to the combined distribution for all trees in the model [
Boosting is one of several classic methods for creating ensemble models, along with
bagging, random forests, and so forth. In Azure Machine Learning Studio [
Although neural networks are widely known for applications in deep learning and
modeling complex problems, such as image recognition, they are easily adapted to
regression problems. Any class of statistical models can be termed a neural network if
they use adaptive weights and can approximate non-linear functions of their inputs.
Thus, neural network regression is suited to problems where a more traditional
regression model cannot fit a solution [
The layers of a neural network are made of nodes, the place where computation
happens. A node combines input from the data with a set of coefficients, or weights,
that either amplify or dampen that input, thereby assigning significance to inputs with
regard to the task the algorithm is trying to learn. These input-weight products are
summed and then the sum is passed through a node’s so-called activation function, to
determine whether and to what extent that signal should progress further through the
network to affect the ultimate outcome. If the signals pass through, the neuron has been
“activated.” Fig. 2 shows a diagram of what one node might look like [
A node layer is a row of those neuron-like switches that turn on or off as the input is
fed through the net. Each layer’s output is simultaneously the subsequent layer’s input,
starting from an initial input layer receiving your data. Pairing the model’s adjustable
weights with input features is how we assign significance to those features regarding
how the neural network classifies and clusters input [
The case study chosen is part of the Irish pilot from RESPOND project [
The energy consumption data comes from utility’s electricity meter, which is the most common source of this kind of measurement. It does not consider social aspects, such as number of family members, because it would be hard to track changes considering large groups, thus making the generalization process complex.
This stage is where the selection of features for energy production or consumer consumption prediction model happens. For instance, the production prediction model uses of almost all available variables, only energy consumption is removed from the dataset. On the other hand, consumption prediction does not depend on PV production or some weather features, such as wind speed or solar radiation, so the final dataset is reduced in that case. This practice helps the algorithms to converge faster and allows a better generalization. 4.3
This process aims to make the dataset as homogeneous as possible. Data cleaning works identifying incomplete, incorrect, inaccurate or irrelevant parts of data and then replaces, modifies or deletes the dirty or coarse data. The normalization process aims to change the values of numeric columns in the dataset to a common scale, without distorting differences among variables values. For example, temperature values range from 0 to 25, while solar radiation can achieve values greater than 500.
The normalization method applied to the proposed experiment is the Z-Score, where the values in the specified columns are transformed using equation 1. = −
( ) ( ) (1) where mean and standard deviation are computed for each column separately. 4.4
Once data preparation is done and ready for processing, the dataset is randomly divided and follows two different paths: • 70% of data is sent to training models. • 30% of data is separated and used for testing purposes, to be compared against trained models later. 4.5
The training dataset received from previous stage is trained across four different regression methods: Linear, Decision Forest, Boosted decision tree and Neural Network. Detailed description about each one can be found in the Section 3. 4.6
The predicted results from the different models are compared against the real data. Our experiment considers a performance evaluation of daily and hourly predictions for both, PV generation and energy consumption. Azure ML Studio provides outputs about
overall accuracy, but we have also added a Python script to plot and calculate additional outputs allowing an intuitive visual analysis.
Over the performance evaluation stage, the four algorithms have been parameterized in different ways targeting to minimize errors. Azure ML Studio allows us to input a range of parameters to be trained. For instance, instead of using a static value for number of decision trees in the Decision Forest method you can set a range of numbers and the algorithm will try all of them, choosing the best combination of parameters. Immediately below you can find the final parametrization of each method: • Linear Regression: Method: Ordinary Least Squares. Regularization weight: 0.001. • Decision Forest Regression: Resampling method: Bagging. Number of trees: 8.
Maximum depth of the decision trees: 32. Number of random splits per node: 128. • Boosted Decision Tree Regression: Maximum leaves per tree: 20. Minimum number of training instances: 10. Learning rate: 0.1. Total number of trees constructed: 100. • Neural Network regression: Hidden layer specification: Fully connected case.
Number of hidden nodes: 100. Learning rate: 0.0001. Number of learning iterations: 100. Initial learning weights diameter: 0.1. Normalizer: Gaussian. 5.1
The objective of the model presented in this paper is to provide energy forecast to utilities, in order to help them planning their own energy production in a day-ahead base. The weather inputs for PV production relate to one day before the actual day. The model could also use forecast data from two or three days before, but the accuracy will drop. Users’ consumption prediction considers historical trends and also weather forecast. The results from our tests show a mean error below 14% for PV production prediction and below 6% for energy consumption. The mean absolute error (MAE) ranges from 21% to 35% for PV production and from 10% to 16% for energy consumption. The boosted decision tree presents the best performance across the methods, followed by the neural network. Table 1 presents a compilation of results
In some circumstances, the use of mean error can better represent the reality. For instance, if the utility is analyzing the forecast of a huge number of houses, some of them will present positive errors, predicting more energy than necessary, and others will have the opposite effect with negative errors, so this kind of measurement could result in a better balance than absolute values. Overall, the four methods follow the real data trend, as can be seen in Fig. 4 (PV generation prediction) and Fig. 5 (Energy consumption prediction).
Fig. 5. Energy consumption prediction model comparison.
Hourly energy consumption prediction can be hard to perform, mostly because consumers’ behavior can vary throughout the week with no defined pattern. PV production forecasting can also suffer variation and uncertainty, due to sudden weather changes. In order to show the accuracy across the models, we selected the best performer (boosted decision tree) and the worst (linear regression), both presented in Fig. 6 (consumption prediction) and Fig. 7 (PV generation).
The residuals are grouped by hours, across 150 days of data. As expected, it is easier to predict the end of the night and beginning of morning, when house’s activity is lower and there is no solar radiation. Higher pattern changes mean higher errors.
This work has demonstrated the application of distinct machine learning methods applied to PV energy production and energy consumption predictions, achieving 9.99% and 2.68% of mean error respectively, considering the best case (boosted decision tree). The dataset pattern is unique, so different ML methods should be applied in order to find the best one that suits each specific application.
Due to limitations of our dataset, only two houses in Ireland were analyzed. For future work the dataset will be improved adding houses, hence more historical data. Furthermore, other ML techniques could be applied, examples include Support Vector Machine and Multi-Layer Perceptron neural network. Finally, energy production from other renewables sources and storage systems can be included, adding more complexity to the proposed model.
This research work was funded by the European Union under the RESPOND project with Grant agreement No. 768619.