Long Short-Term Memory (LSTM)

The Long Short-Term Memory (LSTM) model is a type of recurrent neural network (RNN) that is particularly well-suited for time series prediction tasks, including consumption and production power predictions, due to its ability to capture dependencies and patterns over time. More details on LSTM

This application is used for predicting consumption and production power values. Here, for power prediction, LSTM models can analyze historical power data to learn patterns and trends that occur over time, such as:

  • Daily and seasonal variations: Consumption power often follows cyclic patterns (e.g. higher usage during the day, lower at night). Solar production power is high during the day and zero during the nights.

  • External Factors: LSTM can incorporate external factors like weather, day of the week, or holidays to improve prediction accuracy.

Training LSTM for power predictions:

  • Input Data (channels address _sum/UnmanagedConsumptionActivePower): Time series data of past consumptions.

  • Pre-processing: Data needs to be scaled and sometimes transformed to remove seasonality or noise.

  • Training: The LSTM is trained on historical data using techniques like backpropagation through time (BPTT), where it learns to minimize the error between predicted and actual consumption.

  • Prediction: Once trained, the model can predict future power consumption for various time steps ahead (e.g., hours, days, or even weeks).

In practice, LSTMs are favored for their ability to learn complex time-related patterns, making them effective in forecasting energy demand patterns that can inform Energy Management System (EMS) for optimized energy distribution and cost strategies.

Note for activating the predictor

This predictor creates a folder named "lstm" in the OpenEMS data directory (defined by openems.data.dir in EdgeApp.bndrun).

Initially, a generic model will be used for predictions, which may not yield optimal results. However, a training process is scheduled to occur every 45 days, during which the models in this directory will be updated. The 45-day interval consists of 30 days for training and 15 days for validation.

As a result of this process, a new model will be trained and will automatically replace the previous one.

Source Code