Artificial Intelligence Approaches for Energies

In recent years, it has been noted that deep learning, machine learning, and artificial intelligence models are growing in popularity when applying big data for energy control and decision-making processes. In addition, many recently presented prediction models have been based on artificial intelligence and outperform conventional methods, resulting in energy data-related predictions. This is because artificial intelligence provides smart and efficient tools for smart energy systems. Therefore, artificial intelligence models are highly essential for the predictive modeling of energy issues with high accuracy and speed. The integration of artificial intelligence models in energy systems will contribute an extensive range of prospective research directions for the energy community. Innovative artificial intelligence solutions can bring robustness, stability, security, and efficiency to the energy community.

This editorial is entitled “Artificial Intelligence Approaches for Energies” and includes ten papers. A brief introduction of the paper concerned with each of the published articles belonging to this editorial is included below.

A recent renewable energy estimating system mixes physical models with AI to assist system grid integration and operation. The contribution by Haupt et al. [1] introduces such a system produced for the Shagaya Renewable Energy Park [1]. The entirely operational KREPS adopts AI in multiple portions of the estimating structure and procedures, both for short-range and multiple days long-range forecasting. These AI approaches work well in both dynamical and physical models.

The contribution by Lee and Tsai [2] introduced cloud-based AI that operates AI operations in the cloud (AIC) and operates air conditioning systems off the home [2]. The AIC can be shifted any time to yield high control performances without shifting the control HW. The cost of air conditioning is therefore cheaper due to its high efficiency. The proposed AIC control has two-fold conditions: steady cooling rate and a stable range of temperature control. The paper states that the air conditioner demonstrated in this work reached eighth among thousands of air conditioners produced in Taiwan.

The contribution by del Real et al. [3] studies the application of AI approaches to develop energy demand estimation. In order to achieve this goal, the authors introduced a blended architecture consisting of a CNN combined ANN, with the principal purpose of considering the virtues of two structures. (1) Regression competency of the ANN, (2) and feature extraction competency of the CNN. The presented work was trained, validated, and tested in a real-world system. Then, this work was applied in a French energy demand forecast using ARPEGE forecasting weather data. The simulation results section demonstrates that the proposed method outperforms the conventional approach RTE. In addition, the proposed work achieved the highest efficiency result when compared with other benchmarks, including ARIMA and other traditional ANN used approaches.

Energy consumption forecasting is a crucial condition for designing the needed infrastructure for assigning and optimizing the workbooks of battery electric buses employed in urban public transportation. The contribution by Pamuła and Pamuła presented a model employing a decreased number of parameters which reveals prepared bus trips, arrival times, bus stop locations and trip conditions [4]. In this work, an AI-based method is studied for estimating the energy consumption of bus lines. The employed AI-based model is capable of interpretating big data. This characteristic endures even when scaling the approach to various sized transport networks. The network was tested by using real data given by bus authorities of the Jaworzno town in Poland. The forecast of energy consumption was assessed with the results acquired using the conventional regression method based on the collected data. Simulation errors did not overrun 7.1% for the set of several thousand bus trips.

The contribution by Fotiadou et al. studied the monitoring application of a wind turbine (WT) [5]. The authors present two models based on deep features learning, called LSTM-SAE, and CNN-SAE, for semi-supervised fault detection in wind CPs. The learnt internal features facilitate classification studies by allocating each future measurement into its suitable operation status. To complete their schemes’ quality, the results were assessed against real-world wind turbine data. In the simulation results section, the authors could validate that both CNN-SAE and LSTM-SAE give reliable classification scores, demonstrating the high detection rate of the fault alarm level. In addition, with the simple modification of these architectures they can be used on various fault detection (anomaly detection) categories on other CPS.

In the contribution by Kiprijanovska et al., the authors propose a scalable system for day-ahead household electrical energy consumption estimation [6]. This system is called HousEEC. The presented estimation approach is based on AI, especially deep residual neural network, and integrates diverse information sources by extracting features. These features can be obtained from, contextual data such as weather or calendar and historical load of the particular household. In addition, the authors studied new domain-specific temporal features that help the system to better model the pattern of energy estimation of the household. The simulation analysis and assessment were conducted on one of the biggest datasets for house electrical energy estimation: Pecan Street, containing almost 4 years of data.

Wind energy is now considered an economical and promising renewable energy source and has drawn more attention in recent years. In the contribution by Mora et al., a comparative framework is presented where a suite of LSTM RNN models to develop the current gaps and limitations of informed WP (wind power) estimation methods are presented [7]. These integrated networks were conducted on an iterative procedure of changing hyper parameters to better validate their work, and the overall achievement of each architecture, when tackling 1–3 h in advance WP estimation. The proposed error analysis model demonstrated low error variability with a higher performance when the networks learn on a weekly basis. From the results, the authors suggest that the LSTM is the best approach in forecasting one-hour ahead WP conditions.

In the contribution by Desportes et al., the authors studied the hybrid energy storage system controller with hydrogen storage and a lead battery [8]. The authors planned to the decrease carbon emissions of buildings over a long period, while guaranteeing that 35% of the building energy consumption would be generated using energy produced on site. To this end, the authors use the deep reinforcement learning (DRL) approach to develop a control policy as a function of the building and of the storage state. In addition, the authors simulate the possible issues associated with decreasing the action spatial dimension to 1. This idea enhances the presented method performance-wise. Overall, the authors proposed a new algorithm entitled DDPGarep, using a DDPG to study the policy.

In the contribution by Inuzuka et al., the authors studied HEV real-world energy management using DRL with connected schemes such as V2V and V2I [9]. Regarding HEV energy management, it is crucial to run the engine effectively to decrease its total energy cost. This work presents a policy model that considers road congestion and plans to study the distribution of power and mode selection of an optimal system in regard to future policy-based RL. In the experimental results section, a traffic environment is setup in a virtual space by IPG CarMaker and a HEV model is tested on Simulink of MATLAB to compute the energy loss while driving on the road. The experiments demonstrate the versatility of the presented method for the test data; additionally, it showed that considering congestion decreased the total loss and improved speed.

Conventional online scheduling methods depend on accurate forecasts, which are hard to achieve due to the increase in uncertain RESs. To alleviate this issue, in the contribution by Ji et al., the authors present an online scheduling approach for energy optimization based on continuous-control DRL in a microgrid [10]. The authors formulate the online scheduling issue as MDP. The goal of this scheme is to decrease the operating loss of the microgrid by taking into account the vagueness of RESs, such as electricity prices, generation, and load demand. To increase the performance of the scheduling approach, a GRU-based AI network is developed to extract uncertainty features temporally and optimal scheduling decisions are performed in an end-to-end manner. In order to optimize the method, PPO is adopted to train/validate the NN-based policy using data. The authors suggest that the proposed scheme does not require estimated data to overcome the vagueness or earlier knowledge.