1. Introduction
The growing demand for energy, combined with the growing concern about environmental pollution and greenhouse gas production, is contributing to the increased use of renewable energy sources, including solar, wind, biomass, hydraulic, and geothermal sources. Although these sources are diverse, each has a unique energy supply issue. This uniqueness is mainly due to the energy source, the load demand to be met, and the implementation conditions. Moreover, the renewable sources that can be used in an area may not meet the projected energy demands and/or are subject to significant fluctuations. Therefore, renewable sources are often combined with generators for compensatory energy production and storage systems for energy storage to ensure the continuity of energy availability. These combinations are deemed hybrid renewable energy systems (HRESs).
Several technologies and approaches have been proposed to ease the problematic decision-making process related to choosing the right system. These are grouped into four subcategories [
1] with complementary objectives. Pre-feasibility analysis tools help engineers in their initial analysis of the suitability of a renewable energy system. Sizing tools help to find the best values for various parameters, such as the number of solar photovoltaic panels to be used to meet energy demand. These tools consider energy demand as an objective and deal with the problem by searching for values that optimize different objective functions. There are also simulation and open architecture tools. Simulation tools, as the name suggests, are based on user-provided specifications, such as the size of the system to implement. The simulation tool then provides the user with a detailed analysis of the system’s behavior, which is supplied as an input to the simulation model. The last category is open-architecture tools, which are the opposite of the other types (particularly simulation tools, which are primarily black boxes that do not allow for structural modifications). As indicated in their name, such tools offer an open possibility to make modifications thanks to their R&D-oriented components.
The key problem of sizing a hybrid renewable energy system is related to the fact that renewable energy sources cannot consistently produce energy at all times of the day and year, which makes it essential to combine renewables with alternative energy sources. It is also important to find the best size for each source, considering factors such as investment costs and the available installation surface area. Even with a combination of different energy sources, in many cases, it is necessary to add a generator to the system (e.g., to supply energy to medical facilities, which need to be kept powered at all times). Considering when the generator will have to use fossil fuel, depending on the available investment cost, it is important to determine the correct generator size to limit this consumption as much as possible to reduce greenhouse gas emissions. In addition, given the variability of renewable energy production sources, it is necessary to attach external systems.
Adopting a hybrid renewable energy system requires carrying out three essential steps.
Step 1: Scenario generation—This step consists of defining different potentially feasible configurations of these systems, taking into account various factors such as the renewable energy sources available (solar, wind, biomass), the different energy demand profiles that the system must meet, and, above all, environmental and economic constraints (e.g., the location and total cost of the system must be taken into account) [
2]. The literature abounds with methods, such as genetic algorithms (GAs) and probabilistic methods (Monte Carlo), which enable the definition of realistic data sets and scenarios based on temporal correlations and uncertainties [
3]. Considering that renewable sources alone do not guarantee system resilience, the generation process can integrate not only storage sources (batteries) but also external production sources (generators) [
4,
5].
Step 2: Simulation of previously generated configurations—Simulation tools and software, such as HOMER/HOMER Pro (version 3.11.6561.20287) [
1,
2,
6], are available to evaluate the performance of these configurations [
7,
8]. A large body of research has used optimization algorithms, such as particle swarm optimization (PSO) [
9,
10] and simulated annealing, with the aim of optimizing implementation costs and the proportion of renewable energy [
11,
12].
Step 3: Feasibility analysis—This final stage is crucial in determining which of the many final systems will be adopted. It consists of evaluating systems in terms of their technical, economic, environmental, and social aspects. In this stage, energy costs, net present value, CO
2 emissions, job creation rates, and many other factors are evaluated to determine the optimal configuration [
13,
14,
15]. Implementing these different steps, though necessary, requires significant costs in terms of both time and resources. These processes can take several tens of hours [
16,
17], depending on the different characteristics considered.
Simulation time is a significant research issue across multiple fields. The problem has attracted the attention of many researchers in various fields, including the energy field [
17,
18,
19,
20,
21] and in the field of image characterization and reconstruction [
22].
Table 1 summarizes the key contributions of previous work on hybrid renewable energy systems. Martinez-Turegano et al. (2019) [
21] have addressed the problem of extended simulation time for wind farms through developing admittance models with the aim of reducing the mathematical complexity and maintaining acceptable accuracy in performance. This approach has proven very useful in the context of large-scale networks, such as offshore wind farms. In light of this work, Banihashemi et al. (2022) [
19] proposed the use of the auto-encoder approach due to its ability to reduce the size of system parameters. In particular, the use of auto-encoders enables the extraction of the essential characteristics of energy systems, considerably reducing simulation times. This approach has many advantages when researchers have sufficient and representative data at their disposal. Regarding the work by Tounsi (2022) [
17], the approach consisted of replacing modules which were deemed to be complex in the simulation model. Although this approach may address the issues of complexity and simulation time, it remains domain-specific. Finally, in a similar vein to the work of Banihashemi et al. and Tounsi, Mange and Skowronska (2023) [
20] have proposed the use of machine learning (ML) models. The aim of their research was to replace the entire simulation model with predictive models. This reduces processing time while guaranteeing uncertainty management. In this approach, the ML models need to be trained on high-quality data sets and their reliability assessed based on robust model validation.
Although these approaches have been effective in their application frameworks, there are a number of limitations. First, reliable data are required to construct machine learning models. In addition, simulations relating to sizing hybrid renewable energy systems remains a problem relative to each situation. This relativity is an essential problem, as load demands are not the same for two different consumers and energy production sources can differ considerably. When considering the use of a simulation model for hybrid renewable energy systems, such as HOMER (Hybrid Optimization Model for Renewable Energy) [
1,
23], it is not possible to develop a replacement component for an entity or for the entire model. In many cases, a researcher only has simulation rights for simulating the candidate systems they want to install. Next, the researcher must select an analysis methodology to adopt for the simulation data after running a simulation. The completion of the simulation(s) can be a very long process [
16,
24].
Various studies have also addressed multi-processor execution techniques [
25,
26]. A key concern regarding multi-processor execution is that it requires a machine with a multi-core processor and high processing speed [
25]. Although many methods have been discussed in the literature, these methods can only be applied in well-defined study cases. Their application in our study is limited by (i) the non-availability of historical data to best represent all zones and (ii) the impossibility of having modification access to the model, in order to plan the feature extraction study or the replacement of a part of the model by a new optimized model block. Therefore, this study proposes a new methodology for reducing the simulation waiting time through applying a hybrid method based on machine learning and a heuristic search with no historical data or modification of the initial model. With the aim of simplifying the evaluation process for hybrid energy systems, the proposed hybrid methodology combines the k-Nearest Neighbors and Branch and Bound methods. This approach reduces the waiting time for simulation and feasibility analysis tools.
The main contributions of this work revolve around the following three points:
Innovative hybrid kNN and Branch and Bound (BB)—the integration of kNN with BB optimization establishes a robust framework for the selection process of energy systems;
Targeted exploration optimization—the application of BB enables a solution to the problem to be found through analyzing nodes and branches. This approach avoids the exploration of unnecessary branches based on the results obtained in the previous steps;
Dynamic definition of sets of hybrid renewable energy systems—the kNN approach defines sets of feasible systems sharing common characteristics, such as similar neighboring systems. This approach allows for a preliminary classification of systems, which speeds up the search process.
The remainder of this article is structured as follows.
Section 2 presents the methodology, beginning with our three-phase model, and describes how the two algorithms (kNN and BB) help to achieve the objective of the study. The results and comparison with traditional methods are presented in
Section 3. The article ends with concluding remarks and recommendations for future work in
Section 4.