We at Infinite Devices GmbH here get our hands into all kinds of stuff. Like anyone else, we want to predict the future, YOUR future! With infinimesh you will soon be able to see what’s not there yet: predictive analytics. for your data. by AI. Our very first AI feature is about to be released. Created by Anjo from our AI team, we give you now even more tools to plan what’s ahead of you and your company. 


Infinimesh AI gives you accurate forecasts you can build on. Behind the scene works an Artificial Neural Network for you, a Long-Short Term Memory (LSTM) Neural Network (NN). It takes your uploaded data and uses the input variables efficiently to predict steps into the future.

LSTMs are part of the Recurrent Neural Network (RNN) family. Unlike standard NNs, they do not forget the information from 23 time steps ago. They are not starting to process each inputted value without any context, but rather keep the information from previous values and make predictions based upon these. So, just like you, reading each word after another in this post, putting together the sense of the whole text. This makes them perfectly suited for time-series analysis.

Use Case: Smart meter energy data

A perfect example for this tool is sensor data as the sensor doesn’t throw the values into empty space. Each value depends on the previous, so hello there, time-series! With only five columns in our energy data set, we get going. Firstly, some cosmetics: creating the time input variables from the sensor timestamp, preparing the values to be inputted in the LSTM.

To evaluate our LSTM performance, we create a baseline algorithm which simply takes the value from period one as the current prediction in period two. In other words, the values are being shifted by one timestep. The goal for any NN is to minimize a loss function. Which function to use is up to the designer. In our case, we use Mean Squared Error. The output for three different 25-timestamp data chunks looks like this:


Our Baseline model was obviously neither a smart forecast, nor a very accurate one: we get a baseline loss of 25.73.

With that at hand, we move on to the interesting stuff: the LSTM. For our very first forecasting feature, we decided on some rather simplistic model: our LSTM has only two layers. However, as sensor data comes in pretty quick, it has to be quick! Therefore, simplicity matters.

To give you a broader idea of what the predictions look like, we took 50 timestamps for the following plot. Here are three chunks of 50 timestamp length and its predicted values:

Whoop whoop! Not bad for our very first forecasting feature! We were able to decrease the loss significantly with help of the LSTM: 10.59 this time. A lot more accurate, only about 41% of the initial baseline loss!

 From now on, Infinimesh knows which way your energy consumption goes before you know it – perfect for anyone interested in saving real money and real energy! Better for the planet, better for your bank account. Plan ahead and concentrate on what’s important: your business and your decisions. With Infinimesh AI. Let’s get started!

Find the whole notebook with data set on our Github: https://github.com/InfiniteDevices/ai

Scalytics AI Platform: https://scalytics.io
Jupyter Notebook: https://jupyter.org

Recommended Posts