Forecasting Methods for Photovoltaic Energy in the Scenario of Battery Energy Storage Systems: A Comprehensive Review

Abstract

The worldwide appeal has increased for the development of new technologies that allow the use of green energy. In this category, photovoltaic energy (PV) stands out, especially with regard to the presentation of forecasting methods of solar irradiance or solar power from photovoltaic generators. The development of battery energy storage systems (BESSs) has been investigated to overcome difficulties in electric grid operation, such as using energy in the peaks of load or economic dispatch. These technologies are often applied in the sense that solar irradiance is used to charge the battery. We present a review of solar forecasting methods used together with a PV-BESS. Despite the hundreds of papers investigating solar irradiation forecasting, only a few present discussions on its use on the PV-BESS set. Therefore, we evaluated 49 papers from scientific databases published over the last six years. We performed a quantitative analysis and reported important aspects found in the papers, such as the error metrics addressed, granularity, and where the data are obtained from. We also describe applications of the BESS, present a critical analysis of the current perspectives, and point out promising future research directions on forecasting approaches in conjunction with PV-BESS.

1. Introduction

The adoption of renewable energy sources is paramount to facing environmental problems regarding traditional fossil fuels [1], such as global warming, climate change, and air pollution [2]. However, society has an unprecedented energy demand due to the increase in world population, economic development, and the growing use of technology resources. In this sense, the energy transition to cleaner sources needs to progress at a faster pace. According to the Intercontinental Panel on Climate Change (IPCC) (https://www.ipcc.ch/report/ar6/syr/ accessed on 10 February 2022), an increase of 1.5 degrees in the temperature of the earth has been reported since the beginning of the industrial era; therefore, the emission of carbon-based fuels, which have been the main source of energy, needs to be replaced in order to reduce the emission of greenhouse gases that contribute to global warming. Solar energy has been highlighted recently because of the reduced costs of technology and the optimization of the photovoltaic (PV) system. Moreover, theoretically, the total summing of solar energy incidence on Earth can adequately fulfill the world’s current and future energetic demands [3,4].

One of the current challenges for the use of solar energy is its intermittent behavior [5,6]. Weather variations affect solar irradiance, and it can drastically decrease electrical production by the PV system. In this sense, the energy supply from the solar irradiance in a given location can even be interrupted because of the nighttime duration and overcast days. Battery energy storage systems (BESSs) emerge as one of the main parts of solar-integrated power systems to deal with the high variation in solar power generation through power smoothing application [7]. One of the functions of the BESS is to store excess energy from solar generation during periods of low demand and high production and to supply energy during periods of low production and high demand [8]. Therefore, the BESS is paramount to enabling current PV systems to connect to the grid through a stable and secure solar energy supply. It can also be employed in several applications, such as the smoothing of the intermittency of solar energy, energy demand management, energy time-shift (arbitrage), peak shaving, and grid assets planning [7].

Due to solar energy’s non-dispatchable and fluctuating characteristics, forecasting systems have been employed to support decisions regarding the grid or electricity market operators [5,6,7]. This volatile behavior of the solar irradiance can critically influence power, voltage, and frequency balances of the electricity networks, impacting operational and buy/sell energy matters. Solar forecasts can be useful for scheduling, congestion estimation, reserves management, the reduction in energy production costs, and pricing in the energy market [6]. Thus, developing accurate forecasting models is paramount for creating solar energy systems that use the BESS.

Several reviews were conducted in related areas such as in machine learning methods for electrical power forecasting [9], hybrid models for solar radiation forecasting [10], and solar energy generation forecasting based on Artificial Neural Networks (ANNs) [11]. Moreover, comparative analyses with several forecasting methods were performed in solar energy generation [12,13] as a means to understand their accuracy in the field. However, to the best of the authors’ knowledge, the contribution concerning systems that employ solar energy generation forecasting methods with the use of BESS still needs to be explored in depth.

In this sense, this review presents a general overview of the main contributions of published papers that employ forecasting models to solar energy estimation in the scenario of BESSs. We present a review of recent studies over the last six years, focusing on different aspects: the model used to perform the forecasts, the main applications of the BESS, and the deployed metrics to evaluate the final results, among others. The relevance of the topic is presented, and thereafter, the most common solar forecasting methods for systems using BESSs are analyzed from the selected papers. Finally, we discuss the topic, pointing out possible fruitful research directions.

The remainder of this paper is organized as follows: Section 2 presents the methodology adopted to select the papers under investigation; Section 3 shows a summary of the quantitative aspects of the set of the papers; Section 4 describes the forecasting models and applications of the PV-BESS set found; Section 5 presents a critical analysis of the main findings of this research as well as future directions and perspectives; and Section 6 presents the conclusions.

2. Method and Data

The peer-reviewed articles included in this study were selected to answer three research questions (RQ):

• RQ1: What are the scenarios in which solar forecasting and BESSs have been used?
• RQ2: To what extent has power demand forecasting been considered in PV systems with BESSs?
• RQ3: What are the fragilities, opportunities, or gaps in the research of solar forecasting in PV systems with BESSs?

A search on three scientific databases (IEEE, Science Direct, and MDPI) was performed using a query where the filters were:

• The search string (“forecasting” OR “prediction”) AND (“solar” OR “photovoltaic”) AND (“BESS” OR “Battery Energy Storage Systems”).
• The cutoff date was 30 September 2022.
• Matches on keywords, title, and abstract.

The query above returned 278 articles. Then, this initial list was refined manually (using abstracts, titles, and keywords) to exclude papers focusing on internal electric vehicle batteries (state of charge forecasting), domestic consumption applications, or chemical reactions (such as electrolysis), thus resulting in 125 articles. Using a more in-depth analysis, the 125 articles of this refined list were reviewed individually. An even narrower filter was then applied (“solar forecasting/prediction” AND “BESS” must be within the scope of the paper, AND “the use of the forecast/prediction of power” OR “radiation” must be included in the study), resulting in a final list of 49 articles. This is the final number of selected papers that are discussed in this study (

Table 1. Number of selected papers in each scientific database.

Database	IEEE	Science Direct	MDPI
Papers analyzed	22	67	36
Papers selected	10	22	17

). In addition, details of the selected papers, including the year of publication, authors, forecasting targets, and applications are given in

Table 2. Details of the papers reviewed.

Year	Reference	Type of Publication	Forecasting Target	Application
2016	Syed and Raahemifar [14]	Journal	PV power	Cost optimization.
2016	Brenna et al. [15]	Journal	irradiation	Microgrid management.
2017	Conte et al. [16]	Conference	PV power	Planning and control of the real-time operation.
2017	Gao et al. [17]	Journal	PV power	Cost-benefit analysis of PV-BESS power plants.
2017	Massidda and Marrow [18]	Journal	electric load	Optimal electric load forecast.
2017	Peng et al. [19]	Conference	PV power	Connexion of the grid with a wind–solar–battery hybrid power system.
2018	Tayab et al. [20]	Conference	irradiation	Sizing and control strategy of BESSs.
2018	Fathima et al. [21]	Journal	irradiation	Study and analysis of the performance of the Vanadium Redox Battery.
2018	Angenendt et al.[22]	Journal	PV power	Comparison of operation strategies for PV battery home storage.
2018	Litjens et al. [23]	Journal	PV power	Assessment of methods in the forecasting of photovoltaic battery systems.
2019	Agathokleous et al. [24]	Conference	irradiation and PV power	Optimal stochastic operation.
2019	Yang et al. [25]	Journal	irradiation	Optimal scheduling of a BESS.
2019	Wang et al. [26]	Journal	PV power	Scheduling problem.
2019	Montoya et al. [27]	Journal	PV power	Economic dispatch.
2019	Barchi et al. [7]	Journal	PV power	Predictive energy control strategy.
2019	Ellahi et al. [28]	Journal	PV power	Optimal economic dispatch to overcome intermittency.
2019	Pamparana et al. [29]	Journal	PV power	Optimal sizing of a solar photovoltaic.
2019	Liang et al. [30]	Journal	PV power	Scheduling scheme for hybrid energy storage systems.
2020	Kromer et al. [31]	Conference	irradiation	Optimization of the net load of a virtual power plant.
2020	Bakhtvar et al. [32]	Conference	PV power	Optimal scheduling for dispatch.
2020	Fatnani et al. [33]	Conference	PV power	Design of a PV-BESS system for charging stations for electric vehicles.
2020	Mahmud et al. [34]	Journal	PV power	Evaluation of the prediction errors in power demand management.
2020	Kotsalos et al. [35]	Journal	PV power	Optimization framework for energy management and scheduling.
2020	Kiptoo et al. [36]	Conference	irradiation and PV power	Microgrid planning and operations.
2020	Zeynali et al. [37]	Journal	irradiation	Reduce electricity procurement cost.
2020	Yang et al. [38]	Journal	PV power	Optimal dispatch.
2020	Pierro et al. [39]	Journal	PV power	Mitigation of power imbalances.
2020	Batiyah et al. [40]	Journal	irradiation	Power management.
2020	Zhang et al. [41]	Journal	PV power	Minimization of the operating cost of BESSs.
2020	Ku and Li [42]	Conference	PV power	Control strategies of the BESS to support the system operation.
2021	Mohandes et al. [43]	Journal	irradiation	Optimum design for operation of hybrid renewable energy plants.
2021	Badigenchala et al. [44]	Conference	load demand and PV power	Determination of the optimal size of BESSs in a microgrid system.
2021	Cha and Joo [45]	Journal	net load and PV power	Short-term load forecasting.
2021	Tayab et al. [46]	Journal	PV power	Microgrid energy management.
2021	Caines et al. [47]	Journal	PV power	Integrating solar and storage on the grid.
2021	Abarazi et al. [48]	Journal	PV power	Reliable and cost-effective planning framework.
2021	Zsiboracs et al. [49]	Journal	PV power	Grid balancing challenges analysis.
2021	Basu [50]	Journal	PV power	Cost optimization for the expansion of an isolated microgrid.
2021	Beltran et al. [51]	Journal	irradiation	Battery size determination.
2021	Yang et al. [52]	Journal	PV power	Impact of forecasting error on battery sizing.
2022	Dukpa and Butrylo [53]	Journal	PV power	Profit maximization of electric vehicle charging stations.
2022	Klansupar and Chaitusaney [54]	Journal	PV power	Optimal sizing of grid-scaled battery considering power generation costs.
2022	Agharazi et al. [55]	Journal	PV power	Forecasting model for building energy management.
2022	Lagos et al. [56]	Journal	PV power and wind speed	Wind speed and power forecasting methods applied to power systems.
2022	Khajeh and Laaksonen [57]	Journal	PV power	Probabilistic forecasting models and their applications in smart grids.
2022	Conte et al. [58]	Journal	PV power	Improving the efficiency of energy management for smart grids.
2022	Lai et al. [59]	Journal	PV power	Power scheduling strategy for BESSs, aiming for profit maximization.
2022	Rosas et al. [60]	Journal	PV power	Charging and discharging of a BESS using energy predictions.
2022	Yang et al. [61]	Journal	irradiation	Forecasting error analysis for compensation in renewable energy systems.

.

3. Quantitative Analysis

In this section, we present a quantitative analysis regarding the information contained in the selected papers focused on the forecasting models used in PV-BESSs. From the 49 papers selected, the list consists of nine conference papers (18%) and 40 peer-reviewed journals (82%).

In Figure 1, the number of selected works are presented by year of publication from 2016 to September 2022. Note that there is an increase in the number of works over the years in the published investigations, showing the growing appeal for the theme.

Figure 2 shows the number of selected papers from countries where the solar time series were measured. There are series from 21 different countries, with Australia highlighted in six papers. One can note the presence of different regions of the world, such as North and South America, Asia, Africa, Oceania, and Europe. However, it is important to mention that in 10 out of the 49 selected papers, the authors did not comment on the location of the time series used in the experiment.

Figure 3 shows the performance metrics that appeared in more than one mapped paper, while

Table 3. Description of the performance metrics, in which $y_t$, ${\widehat{y}}_t$, ${\overline{y}}_t$, $y_{max}$, and $y_{min}$ are the actual, predicted, mean, maximum, and minimum values, respectively.

Name	Abbreviation	Lower Is Better	Equation
Mean Absolute Error	MAE	True	${\displaystyle \frac{1}{N}\sum _{t=1}^N|y_t-{\widehat{y}}_t|}$
Mean Absolute Percentage Error	MAPE	True	${\displaystyle \frac{100}{N}\sum _{t=1}^N\left|\frac{y_t-{\widehat{y}}_t}{y_t}\right|}$
Mean Bias Error	MBE	False	${\displaystyle \frac{1}{N}\sum _{t=1}^N(y_t-{\widehat{y}}_t)}$
Mean Squared Error	MSE	True	${\displaystyle \frac{1}{N}\sum _{t=1}^N{(y_t-{\widehat{y}}_t)}^2}$
Root-Mean-Squared Error	RMSE	True	$\sqrt{{\displaystyle \frac{1}{N}\sum _{t=1}^N{(y_t-{\widehat{y}}_t)}^2}}$
Normalized Mean Squared Error	nRMSE	True	$\frac{\sqrt{{\displaystyle \frac{1}{N}\sum _{t=1}^N{(y_t-{\widehat{y}}_t)}^2}}}{y_{max} - y_{min}}$
R-Quadratic	R2	False	${\displaystyle 1-\frac{{\sum }_{t=1}^N{(y_t-{\widehat{y}}_t)}^2}{{\sum }_{t=1}^N{({\overline{y}}_t-{\widehat{y}}_t)}^2}}$

presents their abbreviation and mathematical expression. The root-mean-squared error (RMSE) is the most common metric observed, with 13 occurrences. It is also interesting to note that the mean squared error (MSE), RMSE, and normalized RMSE (nRMSE) are directly related since they are based on the squared difference between forecasts and measured values [62]. In this case, 22 papers (44%) applied at least one of these metrics. Some papers did not mention any metric to evaluate their predictions.

This research field commonly organizes the time series data in discrete intervals such as minutes, hours, or days. This definition is crucial for the majority of forecast models (statistical or machine learning) because it determines the granularity (interval) of future forecasts (output of the model). In this sense, the objective of the prediction is also dependent of the granularity: while a very-short prediction (for example, 15 min sampling) is used in the real-time operation, the daily forecasting is applied to the planning of the operation in the following days.

Figure 4 presents the number of papers by granularity. This figure shows that there are three commonly applied granularities: hourly, lower than 16 min, or daily. The first is the most common in the mapped papers, with 23 occurrences. Concerning the “lower than 16 min” granularity, we observed distinct quantities: 15 min, totaling five cases; 10 min, totaling three cases; and one minute, totaling three cases. The daily case was in four scenarios.

Another important topic related to the granularity is the forecasting horizon, which consists of the number of steps ahead which the model needs to forecast. In this context, 65% of the papers specify the applied horizon, which is in the range of 24 h to 96 h ahead. Additionally, 51% of the mapped papers also performed a demand forecast.

Finally, the network representation of the most mentioned terms in the selected papers is presented in Figure 5. In this network, the node size is proportional to the number of connections it presents, and the colors are a result of a clustering process. There are at least seven different groups. It is important to highlight that the terms BESS, photovoltaic, generation, energy storage system, forecast, microgrid, neural networks, and operation present the largest nodes.

Regarding the forecasting approaches, it is possible to highlight the following groups:

• Blue: This group is highlighted by the largest node among the terms, BESS, and it is possible to notice that within the group, the second term with the largest size is photovoltaic, showing the strong co-occurrence between these terms.
• Red: The terms generation, forecast, prediction, microgrid, and model appear in this group. Among the possible forecasting approaches, we observe the term ensemble.
• Green: Despite the fact that the largest node in this group is energy management, we clearly observe the autoregressive model, one of the most important linear forecasting approaches in the literature.
• Purple: This group clearly indicates the neural networks as the largest node, one of the most important techniques to deal with time series forecast at present.

Moreover, Figure 5 shows that a variety of forecasting approaches are employed to predict the solar information in PV-BESSs, such as neural networks, ensembles, and autoregressive models.

4. Solar Forecasting Approaches

This section overviews the models used to perform solar irradiance or solar generation forecasting to optimize systems that use BESSs. Note that there is a direct relation between these variables, but most of the studies address only one of them. As mentioned in Section 3, especially in Figure 5, there is a variety regarding the forecasting approaches in the literature. However, it is important to highlight that many investigations use the aforementioned predictions, but most do not mention the technique used to perform such a task.

In relation to solar irradiance forecasting, several applications for the optimization of dispatch schedule address two main stages: (I) finding the optimal battery size in order to reduce operation costs, and (II) forecasting solar irradiation and solar power generation. The extent of applications may vary since some can be performed in large hybrid photovoltaic plants and others in single households and electric vehicles.

4.1. Irradiance Forecasting

Ellahi et al. [28] performed a review of forecasting models (Weibull distribution with and without using neural networks) employed to predict the availability of renewable sources (wind and photovoltaic power potential) in different periods. Furthermore, a review of the optimization of scheduling through PSO was conducted. Khajeh et al. [57] reviewed probabilistic forecasting models and their application in smart grids, presenting a roadmap for decision-making processes in smart grid environments.

Agathokleous et al. [24] employed an optimal stochastic scheduling model in a photovoltaic plant coupled with a BESS. The parameter selection was performed through a two-stage stochastic mixed-integer nonlinear optimization algorithm, and a neural network was employed to forecast market prices. A case study was conducted using data from Nordic electricity markets and photovoltaic production data from Chalmers University. It is not clear how the predictions were performed. The sensitivity analysis of the BESS size and the expected profit showed that the increase in the BESS size could also increase profitability.

Kromer et al. [31] proposed an optimal dispatch controller using a virtual power plant consisting of BESS and consumer loads on a distribution network at Shirley, Massachusetts, USA. Linear regression was used in the forecasts of the time series of energy load and production. The results indicated that integrating predictions of energy load and solar irradiance into dispatch decisions can increase the storage capacity when compared to a non-predictive baseline. The solar irradiance prediction was not directly performed, employing data from the SolarAnywhere forecasting service by Clean Power Research.

The work from Barchi et al. [7] proposes a predictive energy control strategy. The system uses a combination of production and demand forecasting, which are able to reduce and shift the peak consumption of shopping centers equipped with battery energy storage systems (BESSs). The photovoltaic generation is given by a procedure that involves the measured global horizontal irradiance and air temperature. The database is from Italy.

Cha et al. [45] addressed the problem of an increased behind-the-meter solar generation, which can degrade the accuracy in short-term load forecasting. The authors proposed a semi-supervised approach to estimate behind-the-meter capacity using photovoltaic energy generation and a BESS. Probabilistic net load forecast was performed with the incorporation of behind-the-meter photovoltaic generation.

Gao et al. [17] investigated the intermittence of irradiation in a power plant in China. Moreover, the subsidy policy and the electricity pricing mechanism also posed a challenge. In this sense, a photovoltaic BESS was employed in a power plant. The forecasting model to predict solar irradiation was not mentioned.

Klansupar et al. [54] proposed a method to determine the optimal sizing and the optimal schedule of BESS in order to compensate negative impacts of variable renewable energy (increased operating costs of traditional systems). Mixed-integer programming was used to perform the optimization process, and the results indicated distinct periods for a BESS to operate at maximum capacity for each application. In order to analyze future renewable energy generation, a forecasting method was employed, but further details were not given.

In the literature, there are approaches that perform load forecasting by separating the energy load due to consumer habits from the one due to local weather conditions. However, other approaches perform energy load forecasting without separation, and the choice between approaches can be conducted based on the penetration of distributed production. Massidda et al. [18] performed a comparison between approaches, and the results show that the functional dependency of production needs to be a major role in the optimization of the operation.

4.1.1. Linear Forecasting Models

Some works addressed linear models to perform solar irradiation forecasting. Fathima et al. [21] employed an intelligent battery management system to provide energy storage for a wind and solar hybrid power system. The optimization of the battery size is performed through the use of a bat optimization algorithm [63] and the forecasts are achieved by the use of statistical and multiple linear regression.

In the work of Batiyah et al. [40], the optimal management strategy is achieved through a coordination between converter-based generators, maximum power point tracking algorithms, and a BESS unit to ensure power balance. In order to forecast environmental and energy load variables, the autoregressive integrated moving average model (ARIMA) was used. The results indicated that the proposed system could manage the power load and stabilize voltage.

4.1.2. Ensemble Forecasting Model

In Kiptoo et al. [36], the proposed system consists of wind turbines, solar photovoltaic, and a battery energy storage system (BESS). In this sense, the optimal size of the BESS is achieved by using a mixed-integer linear program, and random forests were used to obtain forecasts of solar irradiance and wind speed. The results obtained from the demand scheduling using the forecasts from the random forest model resulted in a cost reduction of 12.41%.

4.1.3. Neural Network-Based Forecasting Models

There are investigations that address machine learning approaches to predict solar irradiation, mainly based on neural networks. Tayab et al. [20] employed a microgrid energy management system from grid-connected photovoltaic and a BESS. The approach comprised a forecasting module and an optimization module. The first is responsible for the prediction of solar irradiation, temperature, and load demand using a multilayer perceptron (MLP). The optimization module performs the scheduling for power generation and load demand using a particle swarm optimization (PSO) algorithm. The experiments were conducted on a microgrid system located in Griffith University, Australia, which indicated the effectiveness of the proposed system.

Likewise, optimal dispatch for load demand and power generation was invetigated by Brenna et al. [15]. The proposed system is based on BESSs in which a strategy of sizing and control for a photovoltaic generation farm is presented. Neural networks based on the Levenberg–Marquardt training algorithm (MLP) are used to forecast solar irradiation and load demand in 1 h ahead and day-ahead markets.

Fatnani et al. [33] employed an optimized BESS integrated with solar PV in a charging station for electric vehicles. In order to determine the optimal cost of the battery, the PSO algorithm was used based on several variables such as parking area capacity and PV generation capacity. Moreover, an MLP neural network was used to perform one day-ahead forecasts of energy generation and load demand.

Zeynali et al. [37] integrated a cost model based on proper analytical battery degradation in a smart home in order to mitigate its electricity cost. MLP neural networks were used to create a stochastic process and to perform solar irradiation forecasting.

Yang et al. [61] employed Fourier analysis to decompose the components of the solar time series in order to improve forecasting accuracy by analyzing seasonality and trend components. In this sense, long seasonality components were excluded to investigate its influence on the battery sizing problem. Forecasts were performed using different models, including persistence, Elman neural network, wavelet neural network, and ARIMA. Note that Elman’s proposal is a recurrent network, in contrast to the feedforward MLP previously mentioned.

The same research group proposed a control strategy for the optimal scheduling of a BESS in a hybrid photovoltaic and wind power plant while satisfying several operational constraints. The optimization problem was formulated as a mixed-integer linear programming (MILP) problem. Four forecasting models were employed to investigate the impacts in the BESS optimization for day-ahead and hour-ahead horizons: persistence, Elman neural network, wavelet neural network, and ARIMA. It was observed that the ARIMA model reduced the total operation cost by 5% within a year of operation when compared to a persistence model [25].

Abazari et al. [48] proposed a planning framework for the electrification of remote areas. In the proposed framework, an intelligent weather forecasting based on adaptive neuro-fuzzy process and on fuzzy c-means clustering is used to estimate the solar radiation, wind speed, and ambient temperature. The optimal sizing problem is addressed using a Multi-Verse Optimizer (MVO) and is compared against other meta-heuristic methods. The results showed that the employment of photovoltaic energy systems, wind turbine generators, BESSs, and diesel engine generators is a cost-effective approach which resulted in a 96.13% reduction in CO₂ emissions in comparison with systems without photovoltaic energy systems.

The minimum battery capacity size of BESSs for large photovoltaic power plants is explored in the work of Beltran et al. [51]. In this work, numerical weather prediction models are used to perform day-ahead forecasts, while multiple deep learning models are used in the intraday horizon. The forecasting models supported the optimized operation of large photovoltaic power plants under distinct European intraday electricity markets with reduced battery sizes.

4.1.4. Ensemble Forecasting Methods

We also found works that address combination models. Mohandes et al. [43] proposed an energy management system for a hybrid renewable energy plant (HREP) augmented with a BESS. The management system is composed of three main components: the joint forecasting of wind speed and solar irradiation based on deep neural networks and multiplicative weights update; an optimization model for determining the BESS sizing and its operation schedule; and a dynamic ramping limit for rolling hourly dispatch with a criterion to reduce deviations from the hour-ahead dispatch. The results show an increase in profitability of 2.53% through the employment of multiplicative weights’ update strategy.

Tayab et al. [46] proposed a microgrid energy management system based on the improvements in the forecasting and optimization modules. The forecasting module is composed of an ensemble of neural networks with different architectures. Moreover, in the optimization schedule module, a grey wolf optimization (GWO) is employed to improve optimal dispatch. The results indicated that the proposed system achieved improved results regarding its scheduling strategy.

4.2. Photovoltaic Generation Forecasting

Considering the studies involving photovoltaic generation forecasting, the following works are highlighted. Some of them did not explicitly detail the forecasting model used to predict PV power. An example is the work from Syed et al., in which the authors used the PV power forecasting as one of the decision variables to minimize the consumption of a grid located in Canada [14]. The framework introduced in this study, however, does not mention the technique used to perform the predictions. The load demand and BESS status are also used as inputs. The target is to optimize the dispatch of the battery during the high price periods.

4.2.1. Nondisclosed Forecasting Methods

Bakhtvar et al. mention the use of a forecasting unit in its optimal scheduling framework for dispatch [32]. Their paper proposes an energy management system architecture to control a BESS integrated to a wind farm and a solar plant. The forecasting unit sends information to a power estimator that feeds the optimal scheduling unit. It is responsible for providing the necessary information to the real-time control unit. However, there is no clear indication on how the solar power is predicted. The proposal is from Oman.

Agharazi et al. [55] presented a framework to control building energy management, an integrated system that includes photovoltaic (PV) generation, BESSs, and building loads. The proposal was to provide smooth solar injections with a reduction in cost and improvements in the system performance, without addressing ramp rate and PV forecast errors in the system. The PV forecasts are used as input to the proposed method and then an energy trim function is employed to minimize the impact of PV forecasting errors. However, no further information about the forecasting model is given.

Basu investigated the cost optimization of the expansion of an isolated microgrid endowed by plug-in electric vehicle stations [50]. The author considered the heat and power demand growth to propose a framework named short-range heat and power generation augmentation planning (HPGAP), which is executed using variations of the particle swarm optimization, evolutionary programming, and differential evolution to solve the raised optimization problem. The case study presents a system containing diesel generators, a small hydropower plant, a solar PV plant, a wind turbine generator, biomass-fuel-fired boilers, and a BESS. The study does not make it clear how the PV power prediction is made.

Ku and Li introduce a control implementation named intelligent energy management system (iEMS) of a BESS. The case study involves an offshore island microgrid with elevated PV penetration [42]. The objective is to perform load peak shaving and valley filling using a BESS to improve the transient stability of the grid. The authors use data from Taiwan, but they did not mention the model used to perform PV power forecasting. The iEMS automatically performs this task.

The authors from [38] proposed an optimal two-stage dispatch strategy of a domestic PV-BESS. They integrated a generation system under the market environment of the time-of-use price strategy and a dynamic price adjustment strategy to deal with the solar forecast error. The idea is to combine the BESS with an optimal PV dispatch strategy, creating a balance among power supply and battery load, performing peak cutting and valley filling. The authors address three typical scenarios of PV power: a sunny day in summer, a rainy day in summer, and a sunny day in winter. However, there is no clear indication on how the predictions are made.

There are works that use the predictions performed by others, which were mainly made available in specialized websites or from national governments. Zsiboracs et al. [49] performed a study on the challenge of grid balancing, considering interactions between electric grids, photovoltaic power generation, energy storage, and power generation forecasting. The authors observed that BESSs are rarely used to increase the precision of PV power generation schedules. Therefore, they stated that it is difficult to make decisions considering the BESS in this kind of system. They based their study on a comparison between two European countries: Belgium and Hungary. They stated that the PV power generation forecasts influence the level of the annual utilization of energy storage systems. It was not mentioned how these predictions were made, but they used data from the Elia Group in Belgium, while the Hungarian Independent Transmission Operator Company Ltd. provided the Hungarian predictions.

Angenendt et al. [22] investigated operation strategies to improve economic dispatch and consequent grid relief by considering a system with residential PV-BESS. The idea was to improve self-consumption, reduce costs, and utilize peak power generation. The authors considered forecast-based operation strategies, including predictions for load and PV power generation.

In the work by Liang et al., the BESS was used for tracking feed-in schedules (FISs) of photovoltaic (PV) systems. However, they observed that one can only confine the feed PV energy in an error range instead of tracking the FISs precisely because of the FISs’ limited capacity. With the purpose of improving FIS tracking precision and increase profitability, the researchers introduced a three-stage scheduling scheme that uses a hybrid energy storage system composed of batteries and electrolyzers. The authors used the PV power predictions while considering the historical operation data of a PV system, using averages of the 24 h ahead and 15 min ahead PV forecast [30].

4.2.2. Physical Forecasting Models

The work from Litjens et al. [23] introduces a predictive control strategy to increase photovoltaic self-consumption, decrease unused energy (curtailment) losses, and improve BESS revenues. It is important to highlight that unused or not exported electricity from PV systems is lost. The authors state that new control strategies for PV-BESSs can contribute to reduce energy flows to the grid. The inputs of the model are the information about the previous-week, previous-day, next-day weather forecasts, and clear-day radiation patterns. The authors highlight the dependence of the system on accurate forecasts of PV electricity production and electricity consumption. Four procedures are used to estimate PV power in a region of the Netherlands: (a) the PV pattern of the previous day; (b) the average PV pattern from the previous week; (c) weather prediction data for the next day; and (d) the use of clear-sky radiation.

A control strategy named novel model prediction control (MPC) is introduced by Zhang et al. in order to minimize the cost of the operation of a BESS during each control step [41]. The numerical weather prediction (NWP) is addressed to achieve solar irradiance and temperature. Then, the PV power forecasting is performed, following a set of steps, and used as input to calculate the BESS charge/discharge power. This process is performed after the optimization of the MPC.

4.2.3. Statistical Distribution Methods

Statistical distributions are also addressed to estimate the PV power production. Peng et al. introduced a smoothing control system named the wind–solar battery hybrid power system (WSBHPS) [19]. The applicability is to reduce the use of a BESS by maximizing the use of solar and wind energy by considering a Chinese case study. The system evaluates the BESS charging state and uses a power fluctuation compensation module. The solar power used as input variable is estimated using the pre-plan power curve embedded on the WSBHPS.

The work from Conte et al. used a Gaussian model to perform the PV power forecasting in day-ahead planning and the control of the real-time operation of a BESS [16]. The data were from Italy. The objective of the study was to maximize the power delivery using the aforementioned technology.

Dukpa et al. [53] developed a framework based on mixed-integer linear programming (MILP) to profit the maximization of a commercial electric vehicle charging station endowed by an off-grid solar photovoltaic energy with a BESS. As input, they addressed solar photovoltaic forecasting and electric vehicle arrival probability by considering a database from Poland. The PV power forecasting is approximated by a beta distribution.

The beta distribution was also addressed in the work of Wang et al. to perform PV power forecasting [26]. The authors used a multi-objective approach based on the Pareto Principle to schedule the optimization of a model that considers wind, photovoltaic power, and a BESS. The goal was to minimize the influence of the uncertainty of wind and photovoltaic power output on the operation of the system.

4.2.4. Neural Network-Based Forecasting Methods

Several works highlight the importance of forecasting models to estimate the future PV power integrated with BESSs. Pamparana et al. developed a framework for the optimal sizing of a solar photovoltaic panel and a BESS, considering a case study centered in Chile [29]. They were interested in integrating solar energy into the operation of a semi-autogenous grinding mill. The solar power prediction methodology was presented in their previous study, in which this task was accomplished via the application of the Markov chain [29].

Mahmud et al. evaluated the impact of prediction errors in domestic energy management systems [34]. The authors modeled a peak demand management system endowed by a BESS, electric vehicles, and photovoltaic systems. The target was to provide demand management to reduce energy cost, which considers, for example, battery lifetime. The forecasting step was performed using an autoregressive moving average model (ARMA) and a feedforward artificial neural network (an MLP).

Montoya et al. developed a method to solve the optimal dispatch problem for a BESS in direct current mode for an operating period of 24 h. The optimization procedure was performed using nonlinear programming. The intention was to provide an economic dispatch in voltage-dependent load models using PV power forecasting as input. These values were predicted by a nonlinear autoregressive exogenous model (NARX) and a neural network (MLP) [27].

Conte et al. proposed an improvement in the efficiency of energy management in a smart grid. The method is based on a stochastic model in predictive control tasks in order to optimize the operations. In this case, the network presents a BESS in its composition. The PV power is used as one of the inputs, which is predicted by a Time Delay Neural Network. A main contribution is the analysis of the relation between forecasting accuracy and economic income [58]. As a case study, they address a public dataset.

The target of the work from Lai et al. is to provide a method to maximize the profit of a BESS, overcoming the negative effects caused by the fast charging of electric vehicles (EVs) [59]. The inputs of the framework are fast EV charging demand, PV power generation, and electricity arbitrage. The Extreme Learning Machines (ELMs) are used to predict PV power, while the Monte Carlo method is used to predict the fast charging demand. The database is from the UK.

Yang et al. investigated the impact of forecasting error on BESS sizing in hybrid renewable power plants that contain solar generation. The study analyzed a combined impact of forecasting errors and battery efficiencies on battery sizing. The authors presented a solution based on a strategy to determine battery energy and power capacities. They listed as a potential economic gain, the reduction in the battery cost if its suitable size were defined. The solar irradiance data were from the global horizontal irradiance, considering a region from Australia. The persistence model, Elman neural network, wavelet neural network, and autoregressive integrated moving average model (ARIMA) were used to perform solar irradiance forecasting. Then, a modeling tool named PV library calculated the PV power, which was used as one of the inputs [52].

Pierro et al. [39] developed a study that provides two strategies for mitigating energy imbalances and related costs arising from the current increased photovoltaic penetration in the Italian grid. Their investigation showed that remote-controlled PV plants can be scaled down and run on cost-optimized BESSs and eliminate the impact of PV penetration imbalance, providing operational security. The authors used an ensemble based on neural networks to predict PV power.

Badigenchala et al. [44] utilized a Long Short-Term Memory Neural Network (LSTM) to forecast PV generation and load demand. The goal was to model battery aging while also considering ambient temperature, state-of-charge (SOC), and C-rate to minimize the degradation cost of the battery.

4.3. Strengths and Weaknesses of Forecasting Models

Physical forecasting methods presented in [23,41] are complex models that could achieve improved accuracy over other models. However, they are modeled according to specific variables inherent to a given location. Due to the fact the physical forecasting models are particularly specific to a given location, they do not present the same flexibility as statistical and machine learning-based models. The latter present the flexibility of being automatically adjusted to different databases.

Data-driven forecasting models can be applied in different scenarios. However, it is important to highlight the strengths and weaknesses of each method. Some works, such as [21,40], employed the statistical linear models Multiple Linear Regression and ARIMA, respectively. Even though these methods are interpretable and less complex, they assume a linear correlation structure in data. Therefore, since real-world data are often composed of a combination of linear and nonlinear patterns, their application could produce inaccurate forecasts.

Nonlinear models such as neural networks presented in several works in the literature [27,34,44,52,58,59] are data-driven approaches that can perform nonlinear mappings in data patterns. They are popular due to the fact that they are known as universal function approximators and could fit well to data obtained in different applications and scenarios. However, the misspecification of parameters can lead to accuracy loss, producing an overfitted or underfitted model [64]. Thus, the parameter selection is of paramount importance to obtain accurate results.

As mentioned above, single models can produce inaccurate results due to their nature (linear or nonlinear) or because of the risk of the misselection of parameters. One way of reducing the aforementioned risk is to employ a combination of forecasting models (ensembles) to improve the system’s overall accuracy. Ensembles have been employed in forecasting systems in the context of the BESS [36,43,46], achieving more accurate forecasts than the ones obtained by single models. The improvement in accuracy is also followed by the increase in the computational cost of training several machine learning models.

5. Discussion on the Research Questions

The research questions (RQs) are addressed as follows.

RQ1: What are the scenarios in which solar forecasting and BESSs have been used?

One observes that solar forecasting in the context of BESSs has been used in a wide variety of scenarios, such as: financial savings, resulting from a more appropriate management of resources in the generation-versus-demand scenario; policy for the use of solar energy, another source of energy (e.g., wind), and BESSs, aiming at financial savings; the definition of BESS charging and discharging routes with attention to peaks in the demand profile; the definition of BESS usage policy based on generation, demand, and tariff policies; the assessment of operation cost and battery life; and the use of forecasting to determine battery size.

Regarding economical benefits that come from the appropriate use of forecasting, one can mention the work of Conte et al. [58], in which the increase in the income by 18.72% is due to exploiting prediction from a forecasting method. It is worth mentioning that the adoption of a PV system associated with a BESS provides benefits that extend far beyond economic results. It provides social as well as environmental benefits in locations where it operates.

RQ2: To what extent has power demand forecasting been considered in PV systems with BESSs?

For systems using solar energy and BESSs, performing both solar forecasting and load forecasting plays a crucial role. Approaches focused on planning with the aim to provide an optimal system operation, which may involve a smart grid, can benefit from the aforementioned forecasts [44]. It is important to mention that any percentage of prediction errors may increase the chance of wrong battery usage policy, which may degrade battery lifetime [34].

RQ3: What are the fragilities, opportunities, or gaps in the research of solar forecasting in PV systems with BESSs?

An important aspect to highlight is the fact that, although a high number of scientific papers on solar forecasting are available in the specialized literature, the percentage of works focused on purposes in systems with BESSs is low, despite the importance and the actuality of BESS usage in electric systems. If the ScienceDirect base is considered, for example, the search for articles in the period from 2018 to 2022 for the string “solar forecasting” has 563 results, while the number of results drops to 53 for the string “solar forecasting” AND (“BESS” OR “Battery Energy Storage System”). If the ScienceDirect base resource is used for the search on “Title, abstract or author-specified keywords”, the corresponding number of results, for the same previous search strings, are 241 and 2.

Hence, the use of forecasting methods focusing on electrical systems that use BESSs is a promising field of research to be explored. In the literature, there are works that report financial savings obtained through the efficient use of solar forecasting (sometimes combined with the use of other renewable energy sources; sometimes combined with load forecasting).

One aspect that cannot be neglected is the high number of works on solar forecasting in the context of electrical systems with BESSs that do not provide readers with enough information to allow the study to be reproducible. Commonly, works do not present essential information regarding time series forecasting, such as: the database used, the location the data were obtained from, the granularity of the series, and the forecasting horizon.

The use of exogenous variables, despite the well-known corresponding benefits for forecasting accuracy, seems to be still little explored or simply not reported by the authors. Regarding the estimation of the forecasting models, some works can attain higher performance by replacing synthetic time series with actual datasets in the training step (data-driven approaches by using, for instance, machine learning algorithms).

An interesting aspect to be considered as a research direction is the planning or operation of electrical systems that use BESS, considering forecasting models in the diverse subtasks. For example, predictive models can be regarded, as in the following tasks: generation (in which one or multiple energy sources can be considered) forecasting, load forecasting, and electricity price forecasting.

Beyond these perspectives, other research fields are involved in using and implementing BESSs. The Circular Economy investigates aspects of the use of objects such as sustainable development, environmental protection, clean production, ecological consumption, waste regeneration and reuse, aligning social equity, environmental quality, and economic well-being [65]. In the same way, Life Cycle Assessment is a method for assessing environmental impacts regarding the life cycle stages of a product, process, or service. In this sense, these methods align with the global sustainable agenda and must be considered when dealing with BESSs. For example, recent investigations have presented the ecological reuse of batteries in recent days [66].

Another perspective is related to the analysis of the total cost of ownership, which can reveal the full and actual costs related to the acquisition of a BESS. This method allows for the understanding and recognition of the costs involved in acquiring raw materials and other inputs, such as product transportation and storage, use, and disposal, including hidden costs that may not be clear [67].

In addition, the investigation of new technologies for storage devices is an extensive field that is in development, such as fuel cells for Hydrogen Energy Storage (HES), Compressed Air Energy Storage, among others [68,69]. Other aspects, such as BESS’s principle, structure, and category have varied repercussions on the design of systems, and further investigations should be presented for those topics. All these points are currently hot topics directly related to the central theme of this investigation, the application, and development of forecasting models, and must be further studied.

6. Conclusions

The increasing adoption of clean sources or the substitution of fossil fuels with renewable sources play a fundamental role in decreasing the environmental impact of the world demand for energy. In energy supply from solar irradiance, Battery Energy Storage Systems (BESSs) emerge as an alternative to face the challenges that come from the intermittent behavior of that source.

In this paper, a review was presented on solar forecasting in the context of BESSs. Forty-nine works published in the last six years were selected to answer three questions: (1) What are the scenarios in which solar forecasting and BESSs have been used? (2) At what extent has power demand forecasting been considered in PV systems with BESSs? (3) What are the fragilities, opportunitiess, or gaps in the research of solar forecasting in PV systems with BESSs?

The forecasting task in the scenario of BESSs is carried out with a variety of purposes, including: the establishment of a relationship between forecasting accuracy and economic income; financial savings; the definition of BESS charging and discharging routes; and the impact of prediction errors in battery lifetime. Forecasting has been accomplished with a variety of techniques, such as autoregressive integrated moving average (ARIMA), random forest, multilayer perceptron (MLP), an ensemble of neural networks, Time Delay Neural Network, Extreme Learning Machines, the persistence model, Elman neural network, and Long Short-Term Memory (LSTM).

In the selected works, it was observed that the root-mean-squared error (RMSE) is the most common metric for assessing the performance of forecasting methods used in BESSs. Three common granularities were observed in the selected papers: hourly, lower than 16 min, or daily.

A promising research direction is the planning or operation of electrical systems with BESSs while adequately using the results of several forecast outputs: generation forecasting, in which one or multiple energy sources can be considered; load forecasting; and electricity price forecasting. The outlined research area presents contributions to improving the planning and operation of electrical systems, particularly those incorporating Battery Energy Storage Systems (BESSs). Integrating various forecast outputs, including generation forecasting, load forecasting, and electricity price forecasting, presents a promising avenue. These forecasts provide essential insights into the dynamic behavior of the energy ecosystem, providing information to aid in decision-making and resource allocation.

Moreover, the evaluation of the impact of regularization techniques, such as dropout and transfer learning, on the performance of solar forecasting techniques based on convolutional neural networks (CNNs), in the context of systems that utilize BESSs, may be addressed. Dropout is an effective approach to enhance the generalization performance of deep learning models, specifically used to address overfitting issues. Transfer learning is another technique that can significantly improve the efficiency of solar forecasting models. By utilizing a pre-trained model as a starting point, transfer learning allows for the reuse of knowledge learned from a different but related task. Furthermore, investigating the impact of regularization techniques such as dropout and transfer learning on solar forecasting models using Convolutional Neural Networks (CNNs) is a notable endeavor. Dropout addresses the challenge of overfitting, aiming to improve the generalization capability of deep learning models. Transfer learning leverages pre-existing knowledge from related tasks. This exploration contributes to the refinement of forecasting accuracy and reliability.

In the context of prosumer models, in which consumers also generate and store their own energy, the accurate forecasting of solar energy generation plays a crucial role. The evaluation of the impact of integrating BESSs with solar forecasting in prosumer models can be performed by reducing the inherent uncertainty in solar forecasting. It becomes possible to assess the consequential impact on scheduling the allocation of energy resources to align with prosumer goals. By coupling BESSs with solar forecasting, the uncertainty inherent in solar predictions can be mitigated. By using that approach, one aims to facilitate energy resource allocation, taking into account prosumer objectives and energy usage optimization.

In recent works, digital twins have been used in energy systems [70,71]. The use of digital twins and their corresponding benefits can be seen as a promising way for the development of systems that use a variety of energy resources, including solar energy and BESS. Recent advancements have spotlighted the use of digital twins in energy systems, offering an avenue for the integration of diverse energy sources like solar energy and BESSs. The incorporation of digital twins can enhance system modeling, monitoring, and management, ultimately leading to more efficient and resilient energy systems.

Another important future direction is to analyze the performance of Explainable Artificial Intelligence (XAI) techniques in systems that use BESSs with solar forecasting in order to enhance the interpretability of established artificial intelligence models. Enhancing the interpretability of AI models can contribute to decision-making.

In summary, this research area holds the potential to improve the way we plan, manage, and optimize electrical systems. By leveraging diverse forecast outputs, exploring advanced techniques like regularization and transfer learning, integrating BESS with solar forecasting, harnessing digital twin technology, and prioritizing explainability through XAI, the energy landscape can become more efficient, sustainable, and adaptable to evolving needs.