How to predict how much to open the gates of a dam to give an optimal service?
Discover our work in the European Brain-IoT project: we develop intelligent behaviors with advanced automatic learning features, to decide how much to drain the Cecebre dam daily.
You can read the original text in English, on the official Brain-IoT project.
Machine Learning in BRAIN-IoT for Critical Infrastructure Management
In BRAIN-IoT, the water management infrastructure of EMALCSA from A Coruna is used as a set of use cases that demonstrate how the platform and tools developed under the project can be used to solve real world scenarios.
One of these use cases has to deal with the operation of the spill gates of the Cecebre dam. Each day a decision must be taken regarding how much the gates shall be opened in order to provide an optimal level of water flow to the river. This decision is based mostly (but not only) on: water precipitation (last days and forecast), current level of volume of water in the dam, and previous days values for water output.
EMALCSA Cecebre dam in A Coruña
This use case was picked as great example to develop a Smart Behaviour with advanced Machine Learning features. EMALCSA provides access to real data from sensors already managed in their SICA platform. The SICA platform is integrated by several EMALCSA facilities as the Cecebre damn, the ETAP from A Telva, the “Casa del Agua” and also the pumping stations and the distribution network regulation deposits.
Additionally, historical data is available providing a large data set of daily measures from the above mentioned parameters for the last 40 years.
Machine Learning (ML) techniques operate in a very different way than classic Artificial Intelligence (AI) algorithms, known as Symbolic AI. In classical programming, human experts design, model and develop a set of rules that produce a “smart” result. In ML, available data as well as the results expected for that data (prediction) is feed to a ML “model” that adapts and learns, becoming the equivalent of a set of rules: but in ML the system is trained and its behaviour could lead to unpredicted results that have to be evaluated.
So far, two different ML methods have been selected in BRAIN-IoT by Improving Metrics to demonstrate the project results in the EMALCSA critical infrastructure use case: Recurrent Neural Networks (RNN) and Autoregressive Integrated Moving Average (ARIMA). A Smart Behaviour named Water Output Flow has been developed and is being used to verify these models.
Improving Metrics is a partner of the BRAIN-IoT consortium in charge of developing machine learning algorithms based on the two selected use cases. The resulting feedback is being used to iteratively enrich both the architecture and modeling tools of BRAIN-IoT in order to support an ambitious subset of classic and state-of-the-art AI/ML technologies using the same set of abstractions. That combined with our model-based approach and security features create the smart behaviours that are central to the BRAIN-IoT approach.
Overview of the ML learning models selected to develop the Water Output Flow smart behaviour
- Recurrent Neural Networks (RNN), a modern Deep Learning technique that performs amazingly well to make predictions in time series data but needs huge quantity of training data. Not surprisingly, a RNN architecture known as Long short-term memory (LSTM) has provided us the best results, measuring the differences between predictions and expected values using root-mean-square-error (RMSE).
- Autoregressive Integrated Moving Average (ARIMA) models, a statistical analysis method that works well for time series data with trend and seasonality. The predictions resulting from the ARIMA model are reasonably valid, not as good as RNN. But ARIMA can calculate these predictions with just a few hundreds of samples whereas RNN models need at least thousands of samples.
Evolution across 35 years for the 3 main variables in the dataset: precipitation level, water lever and water flow
This brief comparison underlines one of the typical problems with general use of state-of-the-art Deep Learning methods: the need of high quantity and quality datasets to train the models. On top of that, the training phase tends to be computationally expensive: the great ability of RNN to learn patterns in ordered sets of data comes at a cost. This presents difficulties in putting models into production environment, something that BRAIN-IoT framework and developing tools will deal too.