Projects:2020s1-1230 Prediction of Grid Power Outages for Home Battery Optimisation
A key feature of having a home battery is that in the event of a grid power outage, the power supply to the home is uninterrupted, until the battery is depleted. This project explores the use of analysing weather data to find the key signals that indicate the likelihood of a power outage and prepare the batteries in the identified areas accordingly. As majority of unscheduled power outages are caused by extreme weather conditions, the effectiveness of predicting weather related power outages and pre-charging the battery is highly desirable to the home battery owner.
Contents
Introduction
Project team
Project students
- Ella Pietraroia
- Callum Gilbert
Supervisors
- Brian Ng
- Kenneth Hee (Tesla)
Motivation
- Residential batteries can provide power to a home during a grid outage.
- These batteries regularly charge and discharge to maximise profits from solar, therefore there is no guarantee it will be charged before an outage.
- Severe weather is one of the leading causes of unscheduled power outages.
- If these outages could be predicted, the batteries in affected areas could pre-charge in anticipation.
- We are proposing the use of advanced machine learning algorithms to accurately model the relationship between severe weather and grid power outages.
Random Forest
A Random Forest model is used as a baseline for comparison to the other models, as it a popular algorithm in previous studies, with mostly consistent results. Random forests are an ensemble of Decision Trees, with a single decision tree being a graph that represents a series of choices made on the features of the data given.
Convolutional Neural Network
CNNs specialise in recognising spatial patterns and are commonly used for image recognition. A filter is designed to recognise a specific feature and is applied systematically to the input grid, scanning every point taking into consideration the elements around it. The input structure of a CNN is multidimensional, hence a number of transformations had to be made to the data to get it in a shape that could be processed by the CNN. First all stations for a given date and time were mapped into their correct spot. These values where interpolated and extrapolated so that an entire grid had the approximated value relative to those original data points. This process was repeated for all the different features, such as temperature, wind speed, etc.
Recurrent Neural Network
RNNs are a type of deep learning neural network that have a feed-back method as well as feed-forward methods within the network. This allows the model to find patterns in sequences of data, such as those in speech recognition. Hence, this network was modified to fit weather and outage data in a time-series format. RNNs preserve state across time, hence need to be fed into the network in sequential order. Each layer of the RNN receives the input for the given time as well as the output of the previous time. As the model predicts further into the future it will retain less of the information from earlier time steps.
Results
The best results from each of the three algorithms provided the results for predicting three or more outages:
- Random Forest: 95.85%
- Convolutional Neural Network: 84.90%
- Recurrent Neural Network: 98.01%
Conclusions
- There are some uncertainties in the dataset, for example outages are manually entered which may lead to information loss, hence diminished results. This could potentially be improved with more consulting with power companies and more data preprocessing.
- The algorithms seem better at predicting clusters of outages compared to singular outages. Hence, the scope changed from a very granular output to something broader.
- In comparison to the other two models the CNN underperformed.
- The RNN has improved accuracy and false positive rate than the Random Forest, but a lowered true positive rate.
- Training times for RNN and Random Forest are much less than CNN, hence these algorithms can be trained on smaller areas at a much quicker rate.
- Future Steps: More data and data preprocessing such as including component data from electricity providers in the model, further research into RNN hyperparameter tuning and ensemble methods, incorporating a temporal element to a Random Forest algorithm.
References
[1] a, b, c, "Simple page", In Proceedings of the Conference of Simpleness, 2010.
[2] ...
