An LSTM (Long Short-Term Memory) is a type of Recurrent neural network, which instead of using neurons like classical neural networks, uses memory blocks. The connections between units in an RNN allow data to flow fowards and backward, thus creating internal memory, which allow it to exhibit dynamic temporal behaviour when processing sequences. This makes them applicable to tasks such as time series modelling.
It can be difficult to train standard RNNs to solve problems that require long-term dependencies due to the vanishing and exploding gradient problem, whereby the network can’t learn the parameters of the model effectively. LSTMs aim to solve this problem by using special units which includes a ‘memory cell’ that can maintain information for a long period of time. It achieves this by using a set of gates that control information entering the memory, when it is outputed, and when the information is forgotten.
While forecasting energy demand using this method on raw consumption data produces competitive results, it can struggle to predict usage between peak hours, which can be unpredictable in terms of timing and absolute values. From a business perspective, mis-forecasting on this time period can be costly!
Solutions we are currently looking into to try and improve model performance include designing our own loss functions to focus the model on desired periods, and Autoencoders. The Autoencoder will act as an automatic feature extracting tool to enrich the raw consumption data.
Jaye Cribb and Zuzana Manhartova