Day‐ahead scheduling of a hybrid renewable energy system based on generation forecasting using a deep‐learning approach

A significant amount of electricity in numerous regions worldwide is used for lighting roads, squares, and other public spaces. Renewable energy can contribute notably to electricity usage for public lighting. This paper focuses on the day‐ahead scheduling of a hybrid renewable energy system (HRES) exploiting solar–wind energy potential to meet the electrical energy needs of public lighting. The studied HRES provides electricity for a Wi‐Fi hotspot and a charging hotspot for the end users and has an energy storage system that ensures a reliable electricity supply without interruptions. The day‐ahead scheduling of the studied HRES is based on electricity generation forecasting by using a deep‐learning approach. Particularly, the long short‐term memory model is utilized, considering the fact that it is able to perceive long‐term dependencies among the time series. Moreover, the model's performance is investigated through the determination of diverse inputs: (a) historical data, (b) weather predictions, and (c) historical data with weather predictions. Multiple scenarios of energy consumption are assumed and applied to optimize the day‐ahead scheduling. A new recommendation method is proposed and applied for day‐ahead scheduling, utilizing the power forecasts to achieve optimum operation and energy savings. The results point out that the utilization of the proposed recommendation method controls loads when a shortage in power generation and battery capacity is forecasted for the day ahead, leading to significant energy savings and minimizing the power demand from the grid.


| INTRODUCTION
A large part of the global electrical energy production mix derives from conventional energy sources. In recent years, the debate over energy has revolved around renewable energy sources (RES). European Union (EU) has set targets and formulated energy policies to address significant energy challenges. 1,2 The Union's dependence on energy imports, rising global demand and shortages of fuels such as crude oil, which contribute to higher prices, global warming, and pollution are included in these challenges. The focus of energy policy in Europe has been shifting toward RES and has been covering issues ranging from technology development to mass production and distribution, from small-to large-scale systems, integrating local and more remote sources, and transforming them from subsidized to competitive.
Renewable energy is omnipresent, free, and abundant, and renewable energy technologies are improving leading to more effective, sustainable, and almost pollution-free renewable energy systems. RES are stochastic in nature, which is the main disadvantage of the employment of renewable energy systems. The risk of oversized components and high operational and maintenance costs could arise from using one kind of RES for energy production. 3 A solution to this, and a common proposed system for energy production in literature, is a hybrid renewable energy system (HRES). There are two or more renewable energy technologies used in an HRES for energy generation to minimize the drawbacks of each one of them.
Wind and solar energy sources are commonly used RES. A basic disadvantage of solar energy is that it is impossible to generate electrical energy at night, in the rainy season, and in the cloudy season. 4 Therefore, wind energy production systems are combined with solar energy production systems in an HRES to deal with this shortcoming. HRES could be combined with a storage system with a capacity large enough to handle the energy demand and provide autonomy. 5 A significant amount of electricity demand in many regions globally is used for public electricity demand for the lighting of streets, squares, and other public spaces. Specifically, street lighting amounts to 15%-40% of the total electricity spent in standard cities worldwide. RES can contribute in a critical way to addressing the electricity demand for public lighting. For the deduction of the electricity demand, the costs for street lighting and the reduction of greenhouse gas emissions, a proposed solution is to set the bulb into dimming mode. To this path, a great amount of research has been done so far. In Wadi et al., 6 a fuzzy controller is designed and proposed for street lighting to manage the level of light's dimming and a case study in Istanbul-Turkey is applied to verify the proposed approach. Also, an intelligent wireless street lighting system is proposed using ZigBee wireless technology to control and manage the light of the street as proposed by Leccese and Leonowicz. 7 Shaneh et al. 8 presented the feasibility of different cases of a hybrid system to supply electricity to highways and street lighting using low-power lamps for Tabriz city in Iran, while Nyemba et al. 9 presented an optimization of the design an HRES of solar and wind energy to power a 160 W streetlight. Moreover, Reddy and colleagues [10][11][12][13][14][15] are dealing with management and scheduling issues in HRES taking into account economic and technical parameters.
Going further, the development of prediction models for the electricity generated by solar and wind technologies could contribute to the day-ahead scheduling and optimum management of the HRES. As far as forecasting is concerned, various photovoltaic (PV) power forecasting methods exist in literature [16][17][18] mainly using historical data. In addition, there are many papers in which predictions were done utilizing neural networks to model the wind fluid dynamics. [19][20][21][22] Regarding the different methods applied to forecast the power, a deviation between the forecasted power and the real one occures. 23 In many papers, numerous day-ahead models have been proposed, based on artificial intelligence techniques, that is, machine learning (ML) and deep-learning (DL) methods. Particularly, in Dewangan et al., 24 the performance of well-established ML and DL techniques has been examined, such as linear regression interactions, support vector regression (SVR) with different kernel functions, feed-forward neural network, and long short-term memory (LSTM). As a result, it was underlined that SVR and LSTM provide a greater accuracy in forecasts. Additionally, from the outcomes of Li et al., 25 the DL models and more precisely the LSTM model, as the forecasting horizon expands, is a more applicable forecaster. This is concluded by Das and colleagues 26,27 and Delgado and Fahim 28 as well.
The present paper targets the day-ahead scheduling of an HRES based on power generation forecasts using a DL approach. Specifically, the LSTM model is chosen, which is a type of recurrent neural network included in the field of DL. Unlike other studies, this research introduces climate prediction time series to the proposed power generation forecasting model. Ten different scenarios have been put together, five for PV power forecasting and five for wind power forecasting, to investigate how the different variables influence forecasting accuracy through the use of diverse inputs. The results point out that the use of climate predictions leads to more accurate power forecasts. Additionally, to optimally manage the HRES energy equilibrium, multiple scenarios of energy consumption are assumed and applied. A new recommendation method is proposed, utilizing the most accurate power forecasts and the assumed energy consumption scenarios as inputs to schedule the day-ahead operation of the HRES. The dayahead operation scheduling application in real-time data of a pilot HRES underlines that a significant amount of energy savings can be achieved using the proposed method.
Briefly, the contributions of this study are: -LSTM model is evaluated for power forecast, for both PV and wind power generation, utilizing both historical and climate prediction data. -A new recommendation method is proposed and applied for day-ahead scheduling, utilizing the power forecasts and achieving optimum operation and energy savings.
The remainder of the paper is organized as follows: Section 2 presents the methodology developed for the proposed day-ahead scheduling. Section 3 includes the power forecasting results, and Section 4 the recommendation method and algorithm. In Section 5, the results of the application of the HRES day-ahead scheduling are described. Finally, in Section 6 the concluding remarks of this study are presented.

| HRES configuration
The HRES consists of an energy production system exploiting solar and wind RES, a storage system and an energy consumption system combining loads of street lighting, Wi-Fi hotspot, and charging hotspots. The configuration of the hybrid system is presented in Figure 1.
A unit using solar energy (PV system) and a unit using wind energy (wind turbine system) compose the electrical energy production system, complemented by an array of batteries that represents the storage system. The electrical energy consumption system includes a lighting system, a WiFi hotspot, and a charging hotspot for the passing by end users. The output electrical power of the electrical energy production system is channeled to the energy consumption system, while the surplus is stored in the storage system. When the produced energy is not adequate for the load needs of the consumption system, the stored electricity is used to cover the energy shortage.

| Data description
The power generation forecast model utilizes both historical and climate prediction data, namely: PV power, solar irradiation, temperature, wind power, wind speed, wind direction (historical variables), and solar irradiation, temperature, cloud cover, wind speed, wind direction (climate predictions), respectively.
PV power generation data, wind power generation data, light-emitting diode (LED) light load data, and WiFi hotspot energy consumption data were collected from a pilot HRES located in Athens, Greece. The PV system's nominal capacity is 300 W and the wind turbine system's nominal capacity is 600 W. The LED light consumes 100 W/h of operation and the WiFi hotpot consumes 5 W/h of operation. The recording frequency of the data is hourly, and the collected HRES data are for the period May 25, 2020-May 24, 2021. For the purposes of this study, a meteorological station in Athens, Greece has been installed, which records the data of the climate variables of solar irradiance, ambient temperature, wind speed, and wind direction for the same period.
The complete feature list in the raw data is summarized in the following The wind power (H) variable, denoting the total measured wind turbine production (Wh), is calculated based on a measurement of the wind turbine's voltage and current (wind voltage in volt and wind charge current in ampere).
Using these data, we can formulate a forecasting problem where, given the weather conditions and the aggregate PV and wind energy production for the previous hours, we predict the RES energy production for the next hour.
Regarding the energy consumption data, the 100 W LED light operates from sunset until sunrise every day, and the 5 W WiFi hotspot for data transmission and the end user's networking operates 24 h a day for data transmission. Although there is no available energy consumption data, a charging hotspot to supply electricity to passing by end users for mobile devices charging (such as mobile phones, laptops, etc.) is also part of the HRES. So, it is assumed that the charging hotspot supplies electricity to passing by end users to charge a laptop, and three hourly electricity consumption scenarios were created to include the consumption from the charging hotspot: minimum, average, and maximum, according to the following: As far as climate predictions are concerned, a database of temperature and cloud cover variables was first developed. 29,30 The assessment of solar irradiance prediction is based on total monthly solar irradiance energy density distribution (kWh/m 2 and month) among the month days, using as a parameter the average solar altitude per day. 31 The total solar irradiance is calculated by Equation (1), using the Sachsamanoglou-Makrogiannis 31 method as follows: where sin(β) is the solar height at a certain geographical point and A and B are parameters of solar irradiance, as presented in Table 1. 31 It is assumed that: (a) when cloud cover is up to 30%, A high and B high are used in Equation (1) (1) for mean solar irradiance conditions, and (c) when cloud cover is greater than 60%, A low and B low are used in Equation (1) for low solar irradiance conditions. The values of parameters A and B are reproduced for each time step using linear interpolation. The components and total solar irradiance on an inclined surface are calculated as presented by Erbs et al. 32 The deviation between the predicted and the actual solar irradiance is defined by the following equation: where I p i ( ) is the predicted solar irradiance and I a i ( ) is the actual solar irradiance. Moreover, the solar irradiance prediction hourly error is illustrated in Figure 2. The figure comprises the period from May 25, 2020 to May 24, 2021.
F I G U R E 2 Real and predicted solar irradiance deviation.
F I G U R E 3 Real and predicted wind speed deviation.
Besides solar irradiance, cloud cover prediction and temperature prediction climate variables are gathered in the same database. The data are extracted from two climate prediction websites 29 : for the period May 25, 2020-June 25, 2020, and YR 30 for the period June 26, 2020-May 24, 2021. At 6 a.m. every day, hourly time series of cloud cover prediction and temperature climate prediction variables are collected covering the next 24 h.
Moreover, for the wind turbine system wind speed and wind direction climate prediction variables are gathered in the aforementioned database. The data are extracted from two climate prediction websites 29 : for the period May 25, 2020-June 25, 2020, and YR 30 for the period June 26, 2020-May 24, 2021. At 6 a.m. every day for the next 24 h, hourly time series of wind speed and wind direction climate prediction variables are collected.
The deviation between the predicted and the actual wind speed is defined by the following equation: where V p i ( ) is the predicted wind speed and V a i ( ) is the actual wind speed. Moreover, the wind speed predic-tion hourly error is illustrated in Figure 3. The figure comprises the period from May 25, 2020 to May 24, 2021.

| Evaluation metrics
Forecasting performance evaluation represents a key aspect of the present work. Six widely used performance indices by Pedro and colleagues [33][34][35][36][37] and Liu et al. 38 are utilized for the evaluation of the model: absolute error (AE), mean absolute error (MAE), normalized root mean square error (NRMSE), mean absolute range normalized error (MARNE), and -R 2 .
These evaluation metrics are defined by the following equations: where P a demonstrates the real generated power, P f is the forecasted power, N is the total number of the observations of the test set, and P a is the average actual power production of the system.
T A B L E 3 Inputs per case for wind power forecast.

Cases Historical data
Climate predictions

Wind speed
Wind direction F I G U R E 6 Photovoltaic (PV) power actual and forecasted curves per cluster.

| LSTM model
The vanilla LSTM model is selected for power forecasting in the present study. LSTM has the ability to detect dependencies between different time series and to provide forecasts based on the information of previous time steps. It is a recurrent neural network and has been widely used in time series forecasting applications. LSTM's aforementioned ability derives from the structure of the unit, consisting of four layers, which is able to control the information flows from the current unit to the next ones. 39 An autocorrelation analysis is conducted on the PV power data, to define the optimal number of previous time steps to be used as inputs to the model, as the LSTM deals with time-series forecasting. Figure 4 presents the autocorrelation curve of six sequential days. When the forecast of the power generation at time t + 1 of the present day is requested, it is obvious that the most correlated values are the (t + 1) − 24 and (t + 1) − 25; for instance, the PV power generation at time t + 1 and t of the previous day.
For the forecasts of the current study, the power production forecasts should be provided at 06:00 of the current day for the next 24 h. This influences the time step selection for the forecasting process. For power forecasts from 07:00 to 23:00 of the current day (d) the recorded data of the previous day (d − 1) can be used. On the other hand, forecasts referring to time from 00:00 until 06:00 of the next day (d + 1) derive from the data of the current day (d). Figure 5 illustrates the proper selection of the lags considering the forecasting hour.
The same autocorrelation analysis has been performed on the wind power data with the same results as presented above, for the optimal number of previous time steps to be used as inputs to the model.

| Examined scenarios
Ten different cases have been formulated to investigate the contribution of climate predictions in PV and wind power forecasting applications, five for PV forecasting, and five for wind forecasting. The current day and the hour we attempt to predict are designated as D and t, respectively. Likewise, variables V and F denote the historical variables (PV power, solar irradiation, temperature, wind power, wind speed, and wind direction) and the climate predictions (solar irradiation, temperature, cloud cover, wind speed, and wind direction), respectively.
For the PV power forecast, in Case#Base the training of the forecasting model is based only on historical values of PV power production and meteorological data (solar irradiation and temperature). Precisely, to forecast the power generation at time t of the current day (D), we use as inputs the PV power, the solar irradiation and  | 1695 the temperature of times t and t − 1 of the previous day, that is, However, for the forecasts referring to hours 00:00-06:00 of day D + 1, the available data of the current day D are used as inputs, that is, Otherwise, in Case#1a and Case#1b, only climate predictions are utilized for the training. The basic difference between them is that in Case#1a, solar irradiation, temperature, and cloud cover prediction data are utilized, whereas in Case#1b, cloud cover prediction time series is exempted from the training process. In both cases, to forecast the PV power at time t of the current day (D), the climate predictions referring to time t of the previous day (D − 1) are used as inputs, that is, F t D − 1 , and for the predictions referring to hours 00:00-06:00 of day D + 1, the available data of the current day D are used as inputs, that is, Lastly, Case#2a and Case#2b are a combination of Case#Base with Case#1a and Case#1b accordingly. Both historical values (PV power, solar irradiation, and temperature) and climate predictions (solar irradiation, temperature, and cloud cover) are used for the training process. The climate prediction variables used in these cases are the same as in Case#1a and Case#1b, respectively. Table 2 demonstrates the inputs for each case.
For the wind power forecast, in Case#Base, the training of the forecasting model is based only on historical values of wind power production and meteorological data (wind speed and wind direction). Specifically, to forecast the power generation at time t of the current day (D), we use as inputs the wind power, the wind speed and the wind direction of times t and t − 1 of the previous day, that is, for the forecasts referring to hours 00:00-06:00 of day D + 1, the available data of the current day D are used as inputs, that is On the other hand, in Case#1a and Case#1b, only climate predictions are utilized for the training. The basic difference between them is that in Case#1a, wind speed and wind direction prediction data are utilized, whereas in Case#1b, wind direction prediction time series is exempted from the training process. In both cases, to forecast the wind power at time t of the current day (D), the climate predictions referring to time t of the previous day (D − 1) are used as inputs, that is, F t D − 1 , and for the predictions referring to hours 00:00-06:00 of day D + 1, the available data of the current day D are used as inputs, that is, Last, Case#2a and Case#2b are a combination of Case#Base with Case#1a and Case#1b accordingly. Both historical values (wind power, wind speed, and wind direction) and climate predictions (wind speed and wind direction) are used for the training process. The climate prediction variables used in these cases are the same as in Case#1a and Case#1b, respectively. Table 3 illustrates the inputs for each case.
F I G U R E 8 Wind power actual and forecasted curves per cluster.

| PV power forecasting results
The PV power forecasting results are included in Table 4. It is clear from Case#1a and Case#2a that when the cloud cover climate prediction data is included in the forecasting process, more accurate predictions could be achieved. Additionally, in Case#2a, where the model is trained both with historical data and climate prediction data, the forecasting accuracy is considerably improved.
Considering the abovementioned, due to the high number of bins and scenarios, the depiction of the real and forecasted PV power curves is quite complex. For decreasing the number of clusters and illustrating the actual and forecasted curves, hierarchical clustering is employed and the ward method is utilized to separate the days according to MAE. The maximum number of clusters is set to three. Case#1b and Case#2b are excluded from the analysis, as in Case#1a and Case#2a forecasts appear more accurate. Taking these into account, in Figure 6 the real and the forecasted curves of each cluster are presented for Case#Base, Case#1a, and Case#2a. Based on Figure 6, Cluster#1 of each case shows the forecasted days with the minimum daily MAE, while Cluster#3 shows the days with the highest daily MAE. It becomes clear that the days belonging to Cluster#1 of Case#1a have the smallest forecasting error compared to Cluster#1 of the other cases. Additionally, Cluster#3 of Case#1a contains the lowest number of days compared to Cluster#3 of Case#Base, and Case#2a. Last, it is noted that in Figure 6, even if the PV curves included in each cluster are presented continuously, they do not represent sequential days.
The scatter plots of real and forecasted power for the cases are presented in Figure 7. From these plots, additional information about the AE between the actual and the forecasted power is extracted, as the plots share a common color map defining the AE. Moreover, the R 2 indicator is presented in every diagram extracted from the linear regression analysis that has been carried out for each case. Based on Figure 7, it can be concluded that the AEs of Case#1a is minimized. Additionally, in Case #1a, the R 2 metric presents the greatest value, which means that the relationship between real and forecasted PV power is stronger.

| Wind power forecasting results
The Wind power forecasting results are included in Table 5. It is clear from Case#1a and Case#2a that when the wind direction climate prediction data is included in the forecasting process, more accurate predictions could be achieved. Additionally, in Case#2a, where the model is trained both with historical data and climate prediction data, the forecasting accuracy is considerably improved.
Considering the abovementioned, due to the high number of bins and scenarios, the depiction of the real and forecasted Wind power curves is quite complex. For decreasing the number of clusters and illustrating the actual and forecasted curves, hierarchical clustering is employed and the ward method is utilized to separate the days according to MAE. The maximum number of clusters is set to three. Case#1b and Case#2b are excluded from the analysis, as in Case#1a and Case#2a forecasts appear more accurate. Taking these into account, in Figure 8 the real and the forecasted curves of each cluster are presented for Case#Base, Case#1a, and Case#2a. Based on Figure 8, Cluster#1 of each case shows the forecasted days with the minimum daily MAE, while Cluster#3 shows the days with the highest daily MAE. It becomes clear that the days belonging to Cluster#1 of Case#2a have the smallest forecasting error F I G U R E 10 Flowchart of the proposed recommendation algorithm. compared to Cluster#1 of the other cases. Additionally, Cluster#3 of Case#2a contains the lowest number of days compared to Cluster#3 of Case#Base and Case#1a. Last, it is noted that in Figure 8, even if the wind curves included in each cluster are presented continuously, they do not represent sequential days.
The scatter plots of real and forecasted power for the cases are presented in Figure 9. From these plots, additional information about the AE between the actual and the forecasted power is extracted, as the plots share a common color map defining the AE. Moreover, the R 2 indicator is presented in every diagram extracted from the linear regression analysis that has been carried out for each case. Based on Figure 9 it can be concluded that the AEs of Case#2a are minimized. Additionally, in Case #2a, the R 2 metric presents the greatest value, which means that the relationship between real and forecasted Wind power is stronger.

| PROPOSED RECOMMENDATION METHOD AND ALGORITHM
The day-ahead scheduling of the HRES would contribute to its optimal operation, and additionally would minimize the electricity needs from the grid for excess load coverage that battery capacity could not provide. The flowchart of the proposed recommendation algorithm is presented in Figure 10. For every hour of the day ahead, data for the forecasted PV and wind power, as described in Section 3, are inserted as inputs. Additionally, led light scheduled operation data from the pilot HRES and charging hotspot and WiFi hotspot power demand scenarios, as described in Section 3, are inserted as inputs. Every day at 06:00 the aforementioned data are provided to the proposed algorithm and for every hour of the next 24 h recommendations are extracted.
After the calculation of forecasted electricity production and consumption of the HRES, all four states of the production-consumption combination are checked as shown in Figure 10. An initial amount of the battery capacity (C bat(t) ) at t 0 = 06:00 is provided by the data set of the pilot HRES as input, while C bat(t) is calculated for every time step by the following equation: where P ch t ( ) demonstrates the charging power for every forecasted hour as presented in the flowchart and P dis t ( ) represents the discharging power for every forecasted hour initially estimated by Equation (10) and afterwards as illustrated in Figure 10: where C bat,min is the minimum capacity of the battery, which is equal to 20% of the rated battery capacity used in the pilot HRES. The outputs of the proposed recommendation method are shown in Table 6. The first load control recommendation, "Set Charging Hotspot OFF" is for stopping the electricity provided by the charging hotspot, which is a secondary load and would not cause significant inconvenience to the end user compared to the street lighting operation. After that, if the load demand still exceeds power production and the battery's discharging power, the second recommendation is suggested, "Set LED Dimming 60%", which is for setting the LED light at 60% of the rated operation mode. Then, if the load demand is yet not covered, the "Set LED Dimming 30%" recommend setting the led light at 30% of the rated operation mode. If these three recommendations would provide no balance in the electricity power equilibrium, using power from the grid is suggested followed by "Set LED ON" and "Set Charging Hotspot ON" recommendations.
Moreover, after calculating the battery capacity at every time step, and if there would occur an excess of power not used to cover loads and the battery would be expected to be fully charged, the surplus power is suggested to be sent to the grid, via net metering. 40 Last, if it is anticipated that there would be no power production and consumption, the recommendation this method provides is that the battery capacity remains the same.

| DAY-AHEAD SCHEDULING RESULTS
The proposed recommendation method has been implemented in the power forecast test results data set as described in Section 3. Moreover, the proposed T A B L E 6 Extracted recommendations.  scheduling strategy has been implemented on sequential days with continuous energy storage capacity cycles. However, to clearly illustrate the results of the proposed method, 10 days, including all four seasons of the year, have been randomly selected to be discussed in the present paper.

| Load results
Load curves with and without recommendations for the selected forecast days are presented in Figure 11. The green area in Figure 11 graphs illustrates the energy consumption reduction as recommended by the proposed method. Specifically, the total load reduction for days in Figure 11A and B is 6% and 14%, respectively. Moreover, for days in Figure 11C, D, E, and F load with recommendations decrease to 47%, 72%, 36%, and 80% accordingly, relating to load without recommendations. Additionally, days in Figure 11G and H appear, respectively, to decrease to 29% and 37% in the total load with recommendations. Last, the total load reduction for a day in Figure 11I is 7%, and that in Figure 11J is 38%. Considering the charts in Figure 11 and the abovementioned presentence, it is observed that days in the summer and spring seasons have less total load reduction and less hours of recommended load reduction than days in the winter and autumn seasons. Also, 2 days appear a total load decrease more than 50%, while others' load reduction is below 50%. Overall, the total load reduction for the 10 selected days is 23%, with a total electrical energy saving of 4.81 kW.

| Battery capacity results
Battery capacity curves with and without recommendations for the abovementioned days are presented in Figure 12. Battery capacity curves have been extracted taking into account real PV and Wind power data and load data with and without recommendations for the selected days. The green and red areas in Figure 12 graphs illustrate the savings in the battery's capacity after the implementation of the recommendations. Furthermore, the red area indicates the period of the day when the HRES would have been connected to use power from the grid, that is, if recommendations would not have been implemented and battery capacity would have failed to cover the needs.
After the implementation of the proposed recommendation method, 6 out of 7 days ( Figure 12C-J), where using power from the grid would have occurred, do not require an additional power supply, while the day in Figure 12J, the grid power supply hours are minimized from 10 to 4. Overall, the total battery capacity savings for the 10 selected days is 80.63 kWh.

| CONCLUSIONS
In this paper, a day-ahead scheduling of an HRES based on power generation forecasting by using a DL approach has been proposed. Specifically, the LSTM model has been implemented, and climate prediction time series have been utilized in the power generation forecasting model. Ten different scenarios have been put together, five for PV power forecasting and five for wind power forecasting, and the influence of different variables on forecasting accuracy by the use of diverse inputs has been investigated. The results point out that the use of climate predictions leads to more accurate power forecasts. Additionally, to optimally manage the HRES energy equilibrium, multiple scenarios of energy consumption have been assumed and applied. A new recommendation method has been proposed, utilizing the most accurate power forecasts and the assumed energy consumption scenarios as inputs to schedule the day-ahead operation of the HRES. The day-ahead operation scheduling application in 10 randomly selected days highlights a total load reduction of 23%, with a total electrical energy saving of 4.81 kW and total battery capacity savings of 80.63 kWh. Finally, further work could be done in the forecasting area by retraining the power forecasting models and creating load forecasting models when data are available from pilot HRESs, as well as in optimizing the recommendation algorithm to achieve autonomous operation of the HRES.

ACKNOWLEDGMENTS
This work is funded by the Partnership Agreement for the Development Framework 2014-2020 between Greece and the European Union in a frame of the action "Research-Create-Innovate" of the operational program "Competitiveness, Entrepreneurship and Innovation" (EPAnEK).