I developed deep-learning based methods for Li-ion battery State-of-Health estimation and solar power generation forecasting.
During my time as a postdoctoral researcher, I also worked in projects related with renewable energies, such as Li-ion batteries and solar power generation forecasting. Below I present the details of each research.
Project Synopsis | |
---|---|
Title: | Application of deep learning for Smart Energy |
Highlights 1: |
| Highlights 2: |
|
Period: | August 2020- Dec 2022 |
Outputs: |
|
Skills & Technologies: |
|
Research details
The term Smart Energy has been used in the last decade to refer to the technologies for designing and identifying the most achievable and affordable strategies to implement coherent future sustainable energy systems [1].
My research in this area has focused on developing deep learning methods that aid in state-of-health estimation of secondary batteries and solar power forecasting.
Deep learning for Sate-of-Health estimation of Lithium batteries
The use of lithium-ion batteries (LiBs) has popularized because of the wide range of potential applications and benefits, such as high specific energy and efficiency, and long life [2]. Battery state is described using several metrics related to the LiBs charge capacity, such as the state-of-charge (SOC), state-of-health (SOH), and remaining useful life.
Recently, battery-state estimation functions can be constructed using data-driven techniques (for example, deep learning methods) for battery-state estimation. For this purpose we considered Electrochemical Impedance Spectroscopy (EIS) data, because it is able to obtain information on different processes that affect the measured system, as shown in Figure 1. At high frequencies (blue shaded area), only fast processes can be observed, whereas at low frequencies (red shaded area), the opposite occurs. Therefore we considered that EIS data is a rich source of information that can be used for SOH estimation of LiBs.

Figure 1. Nyquist plots for real against imaginary impedance showing the EIS spectra of the a Li-ion cell. The shaded colored parts denote different frequencies ranges.
An overview of the experimental method used in this research is shown in Figure 2. EIS data was first collected from batteries under a wide range of frequencies —ν = 0.02 – 20 kHz. Then, the data were preprocessed and transformed in such a manner that they are suitable for use as input to a convolutional autoencoder (CAE). We presented a two-stage neural network prediction model construction that uses a CAE for feature learning as the first step, followed by a 4-layer deep neural network (DNN )for charge capacity estimation as the second step. The proposed network is called CAE-DNN, and using this specialized architecture, the SOH of batteries at different temperatures can be accurately estimated.

Deep learning for photovoltaic power generation forecasting
An important source of renewable energy is the energy obtained from converting sunlight into electricity using photoelectric technology, which is called photovoltaic (PV) solar energy or simply solar energy. However, the use of solar energy has several disadvantages such as the high initial cost of solar panels, the need to a relatively large area of installation, high dependence on technology development and high dependence in geographical conditions. Therefore, data-driven methods can be used for accurately forecasting PV power generation and ensure a stable supply of energy that can satisfy the demand of such energy.
Traditionally, statistical time series forecasting methods — such as ARIMA or exponential smoothing methods— are used for solar power forecasting, however recently recurrent neural networks (RNN) have been used for this purpose. More specifically, Long Short Term Memory (LSTM) networks. However, LSTM networks suffer from the problem of vanishing gradient.
For this reason, we are developing a transformer network that solves the problem of multi-step day-ahead hourly forecasting of PV power generation. The overall procedure of the proposed architecture is presented in Figure 3. The input of the proposed neural network architecture is the aggregation of several data sources — historical PV power generation, weather observation, weather forecast and solar geometry data. The augmented data input is fed into different variations of Transformer networks as well as LSTM baseline models.
Currently, the general performance of the models is compared using data of two solar power plants in South Korea.

References
[1] Lund, Henrik, Poul Alberg Østergaard, David Connolly, and Brian Vad Mathiesen. “Smart energy and smart energy systems.” Energy 137 (2017): 556-565.
[2] Obregon, Josue, Yu-Ri Han, Chang Won Ho, Devanadane Mouraliraman, Chang Woo Lee, and Jae-Yoon Jung. “Convolutional autoencoder-based SOH estimation of lithium-ion batteries using electrochemical impedance spectroscopy.” Journal of Energy Storage 60 (2023): 106680.