Forecasting demand in the National Electricity Market October 2017
Agenda
Trends in the National Electricity Market ● A review of AEMO’s forecasting methods ● Long short-term memory (LSTM) neural networks ● Demand forecasting with LSTMs ● Final considerations ●
Trends in the National Electricity Market
Forecasting demand for electricity AEMO publishes forecasts for demand from as short as 5 minute intervals to long term horizons of 20 ● years. Long term forecasts of demand provide a basis for efficient network planning and for business ● investment decisions Short term forecasts of demand could help participants in the wholesale energy market make ● strategic bids or manage production to minimise risk / maximise profits. AEMO Long term forecast for strong, weak and neutral economic and Scheduled 30min forecast and spot price for NSW consumer outlooks Source: AEMO website Source: AEMO 2017 Energy Forecasting Insights
Long term forecast performance and implications In recent years, AEMO has made forecasts suggesting demand will continue to rise, which ● contributed to over-investment in network infrastructure. Meanwhile demand has continued to fall as rooftop solar adoption rises, appliances become ● more energy efficient and the manufacturing industry declines. Actual vs Forecast NEM demand Growth in Australian energy consumption 1961 to 2012 Source: AEMO data Source: BREE (2013)
Long term forecast performance and implications The result has been a continued in prices as a result of high network costs spread over a ● smaller consumption base. These recent changes to the dynamics of the energy market has rendered traditional ● forecasting techniques irrelevant and inaccurate, creating considerable future uncertainty Trends in electricity costs and consumption Electricity price index compared with CPI Source: Grattan (2013) Source: ABS (2013)
Short term forecast performance and implications In the shorter term, the network needs to be capable of coping with peak demand. Failure to ● do so can result in blackouts, which can destroy businesses and loss of life in extreme cases. Peak demand (which has also fallen) is typically driven by extreme weather events, triggering ● high air-conditioner (heaters in Tasmania) use at the same time as industry needs. Source: AEMO
An review of AEMO’s forecasting methods
Econometric methods for long term forecasting AEMO has traditionally taken a top-down approach to developing long-term forecasts for key ● energy market measures (operational demand modelled using regression techniques; max and min demand are generated through weather simulations). These models are driven by factors such as economic activity, weather, income (as an indicator for ● air conditioner / heater usage) and electricity prices. Since 2014, they have shifted towards adopting a more bottom-up approach by including more ● granular data feeds (appliances, meters) beyond traditional grid measure such as bulk transmission data These models are segmented into residential and business with separate forecasts produced for ● shorter and longer term horizons
Neural networks for short term forecasting For their short term forecast of demand (5 minute), AEMO uses a neural network which was ● developed in 2014 (leveraging a 1997 model) It was chosen over other time series models (linear, moving average and spectral) for its ● ability relatively simply to incorporate nonlinear functional relationships The model uses the difference in the natural log of lags 5 time periods before, and 5 time ● periods a week ago Schematic diagram of load prediction neural network Predicted (X) vs Actual (Y) % change in demand for NSW Source: AEMO Source: AEMO
Neural networks for short term forecasting A December 2016 report prepared for the Australian Energy Market Commission by the ● University of Wollongong found that the neural network model was severely limited in capability: The model does not cope with contextual situations such as volatility, price responses ○ or spikes in demand Advances in machine learning techniques since its development could provide a better ○ alternative (current implementation is based off a 20 year old neural network) The report looks to experience in Germany where a neural network has been used for a 15 ● minute dispatch forecast Weather was found to be a useful inclusion for regional forecasts ○ Inclusion of smart meter data also improved predictive accuracy ○ Seasonality played a significant role in prediction accuracy (more so than climate) ○ The strongest recommendation is to use a ensemble technique made up of a self-organising ● map (unsupervised learning) with a recurrent neural network to reduce the residual error of the SOM
Long Short-Term Memory Neural Networks
Recurrent neural networks (RNNs) RNNs are neural networks with loops, which ● distinguishes them from other neural network designs RNNs can be thought of as chain of the same ● network passing information to the next link. This chained structure lends itself to many ● sequence based applications, such as: ○ Speech recognition (Google Translate, Amazon’s Alexa, Siri), ○ Text generation, Handwritten text recognition and generation, and ○ ○ Image captioning
Long Short-Term Memory (LSTM) networks One of the shortcomings of RNNs is their inability to connect relevant information to the ● point when it’s needed as the gap between these two points grows. “The car arrived and waited outside. “ Clouds are in the predict[sky]” After ten minutes the driver turned off its predict[engine] ” LSTMs specifically address this issue of long term dependencies by changing the structure of ● its repeating layers to include a stream of information to flow through the chain. They also contain contextual long-term or short-term memory cells, that allow them to make ● predictions on historical context, in addition to recency. A series of gate structures help the network to learn when to remember information, and when ● to forget.
LSTMs using python, Tensorflow and Keras Tensorflow was released in early 2017 by Google as an open source machine learning library ● that can be used to build neural networks Released in 2015, Keras is higher level neural networks package for python ● It makes the assembly of neural networks easier, by creating a simplified interface for ● interacting with Tensorflow (and various other libraries)
Demand forecasting with LSTMs
Developing a LSTM for forecasting demand A LSTM neural network model has been trialled using historical demand and temperature data ● ○ Historical demand by NEM region (5 min observations aggregated to 30 min for NSW) Air temperature by weather station (30 min observations @ Bankstown Airport) ○ The data will be prepared as a series of sequences in multiple dimensions ● Using the 2016 calendar year for model development, create sequences of 3 weeks ○ ○ Normalisation of variables will be done within windows The neural network will have two hidden layers made up of 25 and 10 neurons respectively ● ○ Use 10 epochs (iterations) for this proof of concept The output layer will predict a single value ● ○ A window shifted along the sequence to predict subsequent values
Temperature and Energy Demand The relationship between air temperature and energy demand is well understood. ● Source: AEMO
Temperature and Energy Demand Demand is at its lowest when air temperature is mild (between 17 and 21 ℃ ) and rises when ● temperatures rise and fall. Source: AEMO
Temperature and Energy Demand Behaviours of the population are also evident when demand and temperature are ● segmented by time of day Source: AEMO
Types of time series prediction The sequences below describe the different methods of time series prediction considered ●
Outputs from univariate modelling MSE for each epoch suggests additional training is required ● ○ Perhaps driven by dropout taking place in training set and not the validation set
Outputs from univariate modelling Historical demand was used to predict sequence windows of the next day (48 30min ● periods) Root Mean-Squared Error (RMSE): 1330.37 ●
Outputs from multivariate modelling Historical demand and air temperature was used to predict sequence windows of the ● next day (48 30min periods) RMSE: 1057.36 ●
Summary of model performance The table below compares performance (using RMSE) of univariate and multivariate ● models with respect to the different types of predictions made Prediction Type Univariate Model Multivariate Model Single step ( next 30 minute, for 3 months) 291 279 Full test set (3 months) 1674 1647 Multiple sequences (24 hours, for 3 months) 1330 1057
Next steps Optimise the parameters and structure of the LSTM ● Reframe the model to natively predict desired periods into the future (diagram below) ● Currently the model predicts one, then shifts its window of prediction incrementally forward ○ Collect AEMO forecast data ● Compare LSTM forecast and AEMO forecast with actuals ● ○ Benchmark performance with error measure Consider collecting additional measures (such as smart meter data) ● Consider developing an ensemble of machine learning techniques ●
Final considerations
Recommend
More recommend
