Summer Internships (remote)

Job description

Jungle develops and applies Artificial Intelligence to increase the uptime and performance of renewable energy sources. Built on existing sensors and data streams, the company’s technology enables solar and wind energy owners to squeeze more out of their assets, accelerating the world’s transition to renewable energy sources.

We have six full-time Summer Internships available, which are expected to last 2-3 months. 


1. Data Cleaning and Augmentation for Industrial Times Series

Industrial data is dirty. It is noisy, and the sensors have their own mechanisms to make the measurements. In Industrial data, with millions of points, it becomes imperative to develop techniques that allow a thorough technique that allows evaluating the quality of the data.

In the Machine Learning world, there are two main ways to improve your Artificial Intelligence - data and modelling. 99% of academia in the ML community is on modelling efforts, while the remainder is on improving data. Good quality data might mean a faster convergence, and easier to learn patterns, greatly reducing the required data to train a good model.

The main goals of the internship are the development of data cleaning and data augmentation procedures, as well as preliminary quality estimation of a dataset. We will be looking at some articles, and study how to get a perception of the quality of a dataset, and how valuable it is for modelling, and how the data transformations affect its quality.
If you want to work with digital gold, a real world dataset of industrial data and polish it, then this is the internship for you.


2. Sequential Models for Short-term Wind Power Forecasting
Renewable energies are non dispatchable sources of energy - i.e: you can only produce when there is wind/sun. Advancing renewable energy forecasting technology is one of the main drivers to make them economically viable, by improving maintenance planning, grid optimisation and energy trading. At Jungle we're using machine learning to push the boundaries of energy forecasting.

To forecast wind power a few hours ahead the main data source are the last measurements of the production itself. This makes it an excellent candidate to leverage the recents advances in sequential modelling using neural networks. The scope of this internship will be to implement and compare several sequential and autoregressive neural network architectures such as LSTMs, NBeats and TCNs for short-term wind power forecasting.

If you want to work with state-of-the-art deep learning models and real world datasets while making renewable energies cheaper this is the internship for you!

3. Set functions for multivariate time-series

Sensor data coming from large pieces of machinery such as wind turbines have an enormous potential to create predictive maintenance solutions that help both in guaranteeing that the machine health is as expected as well as checking if peak performance is being reached.

Deep learning models are very powerful in extracting complex and highly non-linear dynamics directly from the operation data of the assets. A challenge that such solutions have is the fact that data does not come clean and pre-processed in a format that most deep learning models require.

This work is to assess the impact and advantages of using set functions with deep learning models. These methods have been recently introduced in the context of time series classification (see here). The goal of the work is to test the effectiveness of these models in a regression paradigm.

4. Spatiotemporal Modelling for Wind Power Forecasting

Renewable energies are non dispatchable sources of energy - i.e: you can only produce when there is wind/sun. Advancing renewable energy forecasting technology is one of the main drivers to make them economically viable, by improving maintenance planning, grid optimisation and energy trading. At Jungle we're using machine learning to push the boundaries of energy forecasting.

The main data source for accurate power forecasting are weather forecasts. These forecasts are generated for several temporal horizons and for several points around the globe. This results in very high dimensionality datasets with spatial and temporal correlations. The scope of this internship is to implement and compare several neural network architectures (CNNs, LSTMs, Transformers) that fully leverage the temporal and spatial correlations in the input data.

If you want to work with state-of-the-art deep learning models and real world datasets while making renewable energies cheaper this is the internship for you!


5. Specialized ML models for wind turbines

State of the art electromechanical industrial machines such as wind turbines are monitored via hundreds of sensors and signals on a continuous basis. Jungle makes use of these readings to model the normal behaviour of such assets to both ensure that the machine is performing as expected from a health point of view as well as from a performance standpoint.

A modern wind turbine is a very complex machine that houses many different components inside. These operate in harmony to ensure that wind power is converted to electrical power. Different components are exposed to different contexts. For example, blades are mostly affected by the environmental conditions and their rotation speed, while components housed inside the turbine's nacelle are much more influenced by the turbine operating conditions.

The main goals of the internship are to research the needs and impacts of using different sensors and contextual windows for the modelling of different sensors. The student will create hypotheses, design experiments and make an inventory of the requirements of modern wind turbines sensors.

6. Transfer Learning for Wind Power Forecasting

Renewable energies are non-dispatchable sources of energy, i.e, you can only produce when there is wind/sun. Advancing renewable energy forecasting technology is one of the main drivers to make them economically viable, by improving maintenance planning, grid optimization and energy trading. At Jungle we're using machine learning to push the boundaries of energy forecasting.

In 10 years the wind energy capacity in Europe will nearly double. New wind farms are being built around the world every day. What do they have in common? No historical production data. This poses a big challenge for training machine learning models for wind power forecasting. The good thing is that all wind farms, new and old, share the same fundamental correlations. The scope of this internship is to test several state-of-the-art modelling techniques using neural networks to pre-train a general wind power forecasting model that can then be fine-tuned for specific wind farms with low amount of data.

If you want to work with state-of-the-art deep learning models and real world datasets while making renewable energies cheaper this is the internship for you!


Requirements

Desired degrees include Aerospace Engineering , Biomedical Engineering , Electrical and Computer Engineering, Information Systems and Computer Engineering, Mathematics and Applications, and Mechanical Engineering.

Desired skills for all six internships include Python and pandas. Additional requirements for each internship follow below:


  • Data Cleaning and Augmentation for Industrial Times Series internship: You should have experience wrangling datasets in python and pay attention to detail.


  • Sequential Models for Short-term Wind Power Forecasting internship: You should have experience implementing deep learning models using pytorch or tensorflow. 


  • Set functions for multivariate time-series internship: You should have advanced knowledge with Pytorch. 


  • Spatiotemporal Modelling for Wind Power Forecasting internship: You should have experience implementing deep learning models using pytorch or tensorflow. 


  • Specialized ML models for wind turbines internship: You should be pragmatic and able to design and plan experiments. Desired skills include PyTorch.


  • Transfer Learning for Wind Power Forecasting internship: You should have experience implementing deep learning models using pytorch or tensorflow.