Non-Intrusive Load Monitoring (NILM) is a promising approach for energy disaggregation, enabling the identification of appliance-specific energy consumption from aggregated electrical signals. This study proposes a fast Sequence-to-Sequence (S2S) architecture that integrates a Convolutional Variational Auto-Encoder (CVAE) and a transformer model for accurate appliance disaggregation. Performing the reconstruction based on the CVAE is the key element for improving the S2S disaggregation model. The framework is validated on a residential dataset (UK-DALE) and a laboratory dataset collected in a university setting, which includes three distinct appliances: a laser cutting machine, a dust and fume extractor, and three 3D printers. Experimental results demonstrate the efficiency and accuracy of the proposed method in diverse environments, proving its potential for real-world applications in NILM.

Accurate State of Health (SOH) estimation is critical for optimizing battery management, extending lifespan, and preventing unexpected failures in lithium-ion batteries. However, existing methods often face challenges related to computational complexity and data requirements, limiting their practicality in real-time, resource-constrained applications. This paper presents a novel data-driven approach for SOH estimation that combines low computational demands with high accuracy. Using Gaussian Process Regression (GPR) and features extracted from both charge and discharge voltage curves, we achieve precise SOH estimation with a Mean Absolute Error (MAE) of 0.06%. The proposed approach is validated using the Oxford Battery dataset, with results showing consistent accuracy across multiple cells and under varying training data sizes. Notably, the model achieves robust performance even with limited training data, completing estimations in just 0.05 seconds, making it highly suitable for real-time applications. This work contributes to the advancement of battery management systems by providing a computationally efficient, accurate, and real-time SOH estimation method.

The non-deterministic home appliances’ behaviour makes aggregated power consumption hard to be explored and to identify individual appliances’ consumption (disaggregation) in residential buildings. This paper presents a deep neural network learning scheme in order to disaggregate a main meter’s aggregated signal into 11 appliances’ signals and estimate their individual power consumption. A 1-Dimensional Convolution Neural Network (1D CNN) and Long Short-Term Memory (LSTM) layers are used together to form a sequence-to-point (S2P) and a sequence-to-sequence (S2S) Multi-Target Regressor (MTR) for learning and recognizing the loads. Our model is fed with the home total real power (P), total reactive power (Q) and total current (I) and outputs the disaggregated real power (P) for each appliance. The model was trained and evaluated on the AMPds2 public dataset which results in a global disaggregation accuracy of 93.27% for the S2P model and 87.79% for the S2S model. The S2P model outperforms the existing methods in terms of disaggregation accuracy and the number of disaggregated appliances (11 appliances instead of 9) on the used database.

An experimental comparative study of data compression algorithms is investigated for enhancing energy efficiency in low-powered smart meters. Data compression is able to reduce the RF communication time. We also propose a new lossless compression algorithm to achieve the best tradeoff between the compression ratio and computational costs. The performance of our proposed Run-Length Binary Encoding (RLBE) algorithm is compared to those obtained with other lossless compression algorithms: Huffman coding, Even-Rodeh, Exponential-Golomb, Lempel-Ziv Welch, Fibonacci coding, and the hybrid Bzip2 algorithm. The energy optimization in data transmission has been achieved under different operating conditions. The performance of each compression algorithms has been verified experimentally with real metering datasets from industrial and domestic cases.

An experimental investigation of three adaptive algorithms for tracking the fundamental frequency in electric transmission grids is reported. An adaptive Prony's method, an adaptive notch filter and an extended Kaiman filter have been presented in this paper. The design principles and the validity of the models have been sketched. For each method, appropriate models are presented and parameter adjustment guidelines are also proposed. The algorithms were developed in simple digital implementations compliant to real-time applications, therefore with low computational burden. Their performance have been evaluated and compared with the zero-crossing technique which serves as a reference. They have been designed especially for estimating frequency changes like small jumps and when the signal is corrupted with noise and other disturbances due to harmonics. The frequency estimators are compared on the basis of precision, transient response, and degree of noise and harmonics immunity.

This paper shows the advantage of using neural network modularity over conventional learning schemes to approximate complex functions. Indeed, it is difficult for artificial neural networks like Kohonen extended maps to converge toward an efficient and adequate solution when the dimensionality of the input and output spaces are high. Associated to an appropriate learning technique, modularity in neural networks is able to overcome the high dimensionality of the input/input space by decomposing it into different intermediate spaces of reduced dimensionality. The decomposition results in independent neural modules. The efficiency of this learning technique will be enlightened with a visual servoing application. In this application, the relationship between the visual features issued from a stereoscopic vision system and the angles of a 5 DOF-robot will be learned and approximated. Simulations have been conducted and clearly show that this complex, nonlinear, and high dimensional function can be learned efficiently with the neural network modularity approach. Moreover, we show through these simulations that neural modules can be re-utilized, thus reducing the convergence time of the learning and the memory requirements.