Due to advanced monitoring technologies including the plug-in of the cyber and physical layers on the Internet, cyber-physical systems are becoming more vulnerable than ever to cyberthreats leading to possible damage of the system. Consequently, many researchers have devoted to studying detection and identification of such threats in order to mitigate their drawbacks. Among used tools, Machine Learning (ML) has become dominant in the field due to many usability characteristics including the blackbox models availability. In this context, this paper is dedicated to the detection of cyberattacks in Smart Grid (SG) networks which uses industrial control systems (ICS), through the integration of ML models assembled on a small scale. More precisely, it therefore aims to study an electric traction substation system used for the railway industry. The main novelty of our contribution lies in the study of the behaviour of more realistic data than the traditional studies previously shown in the state of the art literature by investigating even more realistic types of attacks. It also emulates data analysis and a larger feature space under most commonly used connectivity protocols in today's industry such as S7Comm and Modbus.

The clean energy conversion characteristics of proton exchange membrane fuel cells (PEMFCs) have given rise to many applications, particularly in transportation. Unfortunately, the commercial application of PEMFCs is hampered by the early deterioration and low durability of the cells. In this case, accurate real-time condition monitoring plays an important role in extending the lifespan of PEMFCs through accurate planning of maintenance tasks. Accordingly, among the widely used modeling tools such as model-driven and data-driven, machine learning has received much attention and has been extensively studied in the literature. Small-scale machine learning (SML) and Deep Learning (DL) are subcategories of machine learning that have been exploited so far. In this context and since SML usually contains non-expansive approximators, this study was dedicated to improving its feature representations for better predictions. Therefore, a recurrent expansion experiment was conducted for several rounds to investigate a linear regression model under time series prognosis of PEMFCs. The results revealed that the prediction performance of SML tools under stationary conditions could be further improved.

The increase in electricity theft has become one of the main concerns of power distribution networks. Indeed, electricity theft could not only lead to financial losses, but also leads to reputation damage by reducing the quality of supply. With advanced sensing technologies of metering infrastructures, data collection of electricity consumption enables data-driven methods to emerge in such non-technical loss detections as an alternative to traditional experience-based human-centric approaches. In this context, such fraud prediction problems are generally a thematic of missing patterns, class imbalance, and higher level of cardinality where there are many possibilities that a single feature can assume. Therefore, this article is introduced specifically to solve data representation problem and increase the sparseness between different data classes. As a result, deeper representations than deep learning networks are introduced to repeatedly merge the learning models themselves into a more complex architecture in a sort of recurrent expansion. To verify the effectiveness of the proposed recurrent expansion of deep learning (REDL) approach, a realistic dataset of electricity theft is involved. Consequently, REDL has achieved excellent data mapping results proven by both visualization and numerical metrics and shows the ability of separating different classes with higher performance. Another important REDL feature of outliers correction has been also discovered in this study. Finally, comparison to some recent works also proved superiority of REDL model.

Proton Exchange Membrane Fuel Cells (PEMFCs) are a key challenger for the world’s future clean and renewable energy solution. Yet, fuel cells are susceptible to operating conditions and hydrogen impurities, leading to performance loss over time in service. Hence, performance degradation prediction is gaining attention recently for fuel cell system reliability. In this work, we present a knowledge transfer approach for online voltage drop prediction. A dual-path convolution neural network is proposed to extract linearity and non-linearity from historical data and performs multi-steps ahead prediction with uncertainty quantification. Online voltage prediction is then evaluated with and without knowledge transfer using two different PEMFC datasets. Results indicate that our proposed approach with transfer knowledge can predict the voltage drop accurately with a small uncertainty range compared to the conventional approach.

Digital twins have transformed the industrial world by changing the development phase of a product or the use of equipment. With the digital twin, the object’s evolution data allows us to anticipate and optimize its performance. Healthcare is in the midst of a digital transition towards personalized, predictive, preventive, and participatory medicine. The digital twin is one of the key tools of this change. In this work, DT is proposed for the diagnosis of breast cancer based on breast skin temperature. Research has focused on thermography as a non-invasive scanning solution for breast cancer diagnosis. However, body temperature is influenced by many factors, such as breast anatomy, physiological functions, blood pressure, etc. The proposed DT updates the bio-heat model’s temperature using the data collected by temperature sensors and complementary data from smart devices. Consequently, the proposed DT is personalized using the collected data to reflect the person’s behavior with whom it is connected.

Solar radiation forecasting is an important technology that is necessary to increase the performance, management, and control of modern electrical grids. It allows energy regulators to estimate the near-future output power of solar power plants, and can help to reduce the effects of power fluctuations on the electricity grid, thus increasing the overall efficiency and power quality of those plants [1]. However, the variable nature of solar irradiance poses a challenge in the exploitation of solar energy. In this context, forecasting techniques are now essential to ensure sustainable, reliable, and cost-effective solar energy production [2]. This paper proposes a hybrid machine learning model to forecast Global Horizontal Irradiance (GHI) in the short term (1-hour ahead). The experimental assessment of the model is done on the basis of an experimental dataset of 11 years of hourly GHI measurements from the BSRN Tamanrasset station in Algeria. The general framework of the proposed model is explained in Figure 1, and its main steps are summarized as follows:

Solar energy is a vast and clean resource that can be harnessed with great benefit for humankind. It is still currently difficult, however, to convert it into electricity in an efficient and cost-effective way. One of the ways to produce energy is the use of various focusing technologies that concentrate the direct normal irradiance (DNI) to produce power through highly-efficient modules or conventional turbines. Concentrating technologies have great potential over arid areas, such as Northern Africa. A serious issue is that DNI can vary rapidly under broken-cloud conditions, which complicate its forecasts [1]. In comparison, the global horizontal irradiance (GHI) is much less sensitive to cloudiness. As an alternative to the direct DNI forecasting avenue, a possibility exists to derive the future DNI indirectly by forecasting GHI first, and then use a conventional separation model to derive DNI. In this context, the present study compares four of the most well-known separation models of the literature and evaluates their performance at Tamanrasset, Algeria, when used in combination with a new deep learning machine methodology introduced here to forecast GHI time series for short-term horizons (15-min). The proposed forecast system is composed of two separate blocs. The first bloc seeks to forecast the future value of GHI based on historical time series using the Long Short-Term Memory (LSTM) technique with two different search algorithms. In the second bloc, an appropriate separation (also referred to as “diffuse fraction” or “splitting”) model is implemented to extract the direct component of GHI. LSTMs constitute a category of recurrent neural network (RNN) structure that exhibits an excellent learning and predicting ability for data with time-series sequences [2]. The present study uses and evaluates the performance of two novel and competitive strategies, which both aim at providing accurate short-term GHI forecasts: Unidirectional LSTM (UniLSTM) and Bidirectional LSTM (BiLSTM). In the former case, the signal propagates backward or forward in time, whereas in the latter case the learning algorithm is fed with the GHI data once from beginning to the end and once from end to beginning. One goal of this study is to evaluate the overall advantages and performance of each strategy. Hence, this study aims to validate this new approach of obtaining 15- min DNI forecasts indirectly, using the most appropriate separation model. An important step here is to determine which model is suitable for the arid climate of Tamanrasset, a high-elevation site in southern Algeria where dust storms are frequent. Accordingly, four representative models have been selected here, based on their validation results [3] and popularity: 1) Erbs model [4]; 2) Maxwell’s DISC model [5]; 3) Perez’s DIRINT model [6]; and 4) Engerer2 model [7]. In this contribution, 1-min direct, diffuse and global solar irradiance measurements from the BSRN station of Tamanrasset are first quality-controlled with usual procedures [3, 8] and combined into 15-min sequences over the period 2013–2017. The four separation models are operated with the 15-min GHI forecasts obtained with each LSTM model, then compared to the 15-min measured DNI sequences. Table 1 shows the results obtained by the two forecasting strategies, for the experimental dataset.

Nowadays, solar energy, which is the direct conversion of light into electricity, occupies a very important place among renewable energy resources due to its daily availability in most regions of the globe. Therefore, the wise exploitation of this clean energy will ultimately drive to cover all needed demands [1, 2]. This paper deals with the design of Maximum Power Point Tracking (MPPT) technique for photovoltaic (PV) system using a modified incremental conductance (IncCond) algorithm to extract maximum power from PV module. The considered PV system consists of a PV module, a DC-DC converter and a resistive load. In the literature, it is known that the conventional MPPT algorithms suffer from serious disadvantages such as fluctuations around the MPP and slow tracking during a rapid change in atmospheric conditions. Therefore, in this paper, and attempting to overcome the shortcomings of conventional approach. In this work, a new modified incremental conductance algorithm is proposed to find the Maximum Power Point Tracking (MPPT) of the Photovoltaic System. Simulation tests with different atmospheric conditions are provided to demonstrate the validity and the effectiveness of the proposed algorithm.

This paper describes a new approach for hourly global solar radiation forecasting based on a hybrid artificial neural network technique combining a residual neural network (RESNET) for powerful feature extraction of the most relevant moments of the past, and a long short-term memory (LSTM) technique for efficient projection into the future. Based on 11 years of solar irradiance measurements at Tamanrasset, Algeria, four evaluation metrics are used to demonstrate the efficiency of the proposed method: coefficient of determination (R²), root-mean-square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). These metrics are also used to evaluate the performance of the model in comparison with two existing forecasting models used as benchmark: a particular technique of convolutional neural network (CNN) called 1-dimensional convolutional neural network (1D-CNN) and a conventional LSTM. The present results indicate that the proposed RESNET-LSTM model outperforms the other models in terms of all statistical indicators.

This paper proposes an optimum design of a diesel/PV/wind/battery hybrid renewable energy system (HRES) for rural electrification in a remote district in Tamanrasset, Algeria. In this study, a particle swarm optimization algorithm (PSO) has been proposed to solve a multi-objective optimization problem, which was created by carrying out simultaneously, the cost of energy (COE) minimization while maximizing the reliability of power supply described as the loss of power supply probability (LPSP) and a renewable fraction (RF). The simulation results show that the PV/WT/DG/BT is the best economic configuration with a reasonable annual cost of the optimal system (ACS) which is about 7798.71 $ and the COE equal to 0.79 $/kWh for an LPSP = 0.01%, where the ten households are 0.99 % satisfied by renewable energy sources.

Requirements of users in developing countries differ from those of developed countries. This difference can be seen through wheelchair displacement in infrastructures that don’t meet international standards. However, developing countries are obliged to purchase products from developed countries that don’t necessarily meet all user’s requirements. The modification of these requirements will generate disruption on all the supply chain. This paper proposes a model for optimising the cost of requirement modification on the supply chain and seeks to evaluate the introduction of a new requirement on an existing product/process. This model is adapted to the redesign and development of products, such as wheelchairs, satisfying specific Algerian end-user requirements.

A comparative study between a set of chosen machine learning tools for direct remaining useful life prediction is presented in this work. The main objective of this study is to select the appropriate prediction tool for health estimation of aircraft engines for future uses. The training algorithms are evaluated using “time-varying” data retrieved from Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) developed by NASA. The training and testing processes of each algorithm are carried out under the same circumstances using the similar initial condition and evaluation sets. The results prove that among the studied training tools, Support vector machine (SVM) achieved the best results.

Condition Monitoring of photovoltaic systems plays an important role in maintenance interventions due to its ability to solve problems of loss of energy production revenue. Nowadays, machine learning-based failure diagnosis is becoming increasingly growing as an alternative to various difficult physical-based interpretations and the main pile foundation for condition monitoring. As a result, several methods with different learning paradigms (e.g. deep learning, transfer learning, reinforcement learning, ensemble learning, etc.) have been used to address different condition monitoring issues. Therefore, the aim of this paper is at least, to shed light on the most relevant work that has been done so far in the field of photovoltaic systems machine learning-based condition monitoring.

One of the main data-driven challenges when assessing bearing health is that training and test samples must be drawn from the same probability distribution. Indeed, it is difficult and almost rare to witness such a phenomenon in practical applications due to the constantly changing working conditions of rotating machines. In addition, collecting sufficient deterioration samples from the bearing life cycle is not possible due to the huge memory requirements and processing costs. As a result, accelerated life tests are believed to be the primary alternatives to such a situation. However, and unfortunately, the recorded samples always are subject to lack of real patterns. Therefore, in this paper, a transfer learning approach is performed to solve such kind of problem where PRONOSTICO dataset is used to assess the current procedures.

Nowadays, the integration of solar power into the electrical grids is vital to increase energy efficiency and profitability. Effective usage of the instable solar production of photovoltaic (PV) systems necessitates trustworthy forecasting information. Actually, this addition can gives an ameliorated service quality if the solar radiation variation can be forecasted accurately. In this paper, we propose a new forecasting approach that integrates Autoregressive Moving Average (ARMA) and Genetic algorithms (GA) to make benefit of both of them in order to forecast Global Horizontal Irradiance (GHI) component. The proposed approach is compared with the standard ARMA model. The experimental results show that, the proposed approach outperforms the classical ARMA models in terms of mean absolute percentage error (MAPE), root mean squared error (RMSE) coefficient of determination (R)2 and the normalized mean squared error (NMSE).

In this work a new data-driven approach for Remaining Useful Life estimation of aircraft engines is developed. The proposed approach is a regularized Single Hidden Layer Feedforward Neural network (SLFN) with incremental constructive enhancements. The training rules of this algorithm are inspired form different Extreme Learning Machine (ELM) variants. Particle Swarm Optimization (PSO) algorithm is integrated to enhance tracking ability of the best regularization parameter to reduce the norm of the tuned weights. The proposed approach is evaluated using C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) dataset and compared to its other derivatives and proved its accuracy. C-MAPSS software has revisions in military and civil applications. In this paper, the military version of its application is the used one.

In this work a new data-driven compression approach is presented. The compression algorithm is an autoencoder trained with an improved On-line sequential Extreme Learning Machine (OS-ELM). First, a dynamic adaptation of the training algorithm towards the newly coming data is achieved by integrating an updated selection strategy (USS) and dynamic forgetting function (DDF). Second, Singular Value Decomposition (SVD) is involved to enhance hidden layer representation via sparse mapping. This new developed autoencoder (ASOS- AE) is compared with the ordinary OS-ELM autoencoder (OS-AE) and proved its accuracy in CMAPSS dataset (Commercial Modular Aero-Propulsion System Simulation). The C-MAPSS software has revisions in civil and military applications. In the present work we used the military version of its applications.