Data-Driven Management Systems for Wave-Powered Renewable Energy Communities

Abstract

This research focus on the essential task of precise prediction for power generation and energy consumption of wave energy converters (WECs) within the framework of contemporary wave-powered renewable energy sources (RESs). Utilizing real-time wave data, we introduce a deep learning methodology featuring a long short-term memory (LSTM) model. Additionally, we propose an online management system for RESs aimed at optimizing interactions among WECs, energy storage systems (ESSs), super capacitor (SC), and load. This approach leads to significant enhancements in mean square error (MSE) for critical variables such as wave height, time period, and direction, improving predictive accuracy by factors of 8.37, 9.30, and 16.14, respectively. Through diverse scenario-based experimental evaluations, our solution exhibits competitive performance when compared to benchmark strategies and ideal solutions. These findings underscore the potential of the LSTM-NN model to advance the efficiency and reliability of wave energy forecasting and management systems. As wave energy technology evolves, this study contributes to ongoing efforts to enhance practical applicability, especially in coastal regions with substantial wave energy potential.

1. Introduction

Global power demand, set to reach 30,000 TWh by 2030, is driving the rapid expansion of renewable energy sources (RESs) as part of the solution to combat CO₂ emissions [1]. In 2020, RESs contributed to 28% of global energy demand, with projections indicating it will surpass 45% by 2040, predominantly through solar and wind power [2]. However, wind and solar energy are intermittent, requiring costly grid storage for continuous supply. To improve renewable energy source (RES) performance, various forecasting methods have been explored, including multiobjective optimization algorithms for wind energy [3]. Unlike wind and solar, ocean wave energy offers more predictability and consistency. Yet, its levelised cost of energy (LCOE) remains higher than other RESs [4]. Combining wave energy with other RESs, like wind and solar, can lead to synergistic effects, reducing overall LCOE [5,6].

Wave energy converters (WECs) convert wave mechanical energy into electricity, capturing kinetic energy for various applications. Types include point absorbers [7], oscillating water columns [8], attenuators [9], and overtopping devices [8]. For instance, point absorbers are buoy-like devices that float on water, using hydraulic or pneumatic systems for energy conversion. Eco Wave Power built a 100 KW onshore wave energy converter (WEC) station in Gibraltar in 2014 with plans for 5 MW capacity (Figure 1a). Wave energy converter types are categorized into onshore and offshore units. Development includes installations like WaveRoller and Limpet onshore units and offshore units like Pelamis and AquaBuOY, harnessing wave energy through various mechanisms. Ongoing research is essential for improvement, particularly in coastal regions with limited grid access [10,11,12,13]. Common offshore WEC forms include large floating buoys connected to the ocean floor via tethers (Figure 1b). Efficient power absorption in a WEC array depends on factors like arrangement, spacing, wave height, wave peak time period, and wave directions. Accurate estimation of WEC power output in uncertain offshore weather requires estimating wind speed, wave height, wave peak time period, and wave direction.

Various approaches have been proposed for predicting ocean wave heights, with recent advancements including the minimum radial basis Function (RBF) [14], evolutionary product unit neural networks [15,16], discrete wavelet transform neural networks, and extreme machine learning [17]. In a comparative study [18], authors evaluated the forecasting performance of an artificial neural network (ANN) model against traditional time-series models using major wave heights along the Indian coast. The ANN model demonstrated comparable accuracy to the time-series model for long-term predictions, while exhibiting superior accuracy for short-term predictions. Long-term forecasting of ocean wave time-series is crucial for future grid planning and operation. Notably, there is a lack of prior studies utilizing LSTM-NN neural networks for forecasting ocean wave behaviour, including wave height, peak time period, and wave direction. The application of LSTM-NN in deep waters, where nonlinear dynamic effects are often more pronounced, remains unexplored according to the available literature.

To utilise the wave energy using REC, energy scheduling for REC operation is crucial to achieve both economic and continuous supply objectives [19]. In order to have minimal effect on the stability and robustness performance of the REC, the scheduling function in RECs is performed at an energy management level, which operates at a lower bandwidth than control and power management levels [20]. The control and power management strategies, such as voltage and frequency regulation, can affect the current and voltage quality of the electrical power system. These strategies are designed to maintain the stability and reliability of the system by ensuring that the voltage and frequency remain within acceptable limits [20]. In particular, the power management level is responsible for managing problems such as line constraints, distribution and converter losses, and autonomous operating strategies that can affect the system’s performance [21]. The existing REC energy management system (EMS) frameworks are mainly composed of mathematical formulations, which are validated using deterministic and offline scenarios [22]. A scenario-based day-ahead EMS for ocean RECs is proposed in [23]. The authors in [24] proposes a linear mathematical model for managing the generation and load of a hybrid REC that minimizes the system’s overall operating costs over a 24 h period. In [25], the controller achieves power sharing and energy management by combining fuzzy control with gain scheduling algorithms. In [26], a heuristic method is used to define and solve an economic dispatch problem for the minimization of total operation costs in DC RECs. In [27], a centralized REC-EMS for isolated RECs is developed for optimal power flow to avoid the use of a mixed-integer nonlinear formulation. In [28], a method for optimal scheduling of a combined heat and power system and an RES using a backtracking search algorithm is proposed. The proposed method is effective in reducing operational costs and increasing the use of the RES, as demonstrated through a case study of a hospital building in Tehran, Iran. In [29], an online optimal power control method is suggested for the operation of energy storage in a grid-connected REC, which considers predicted electricity consumption and generation. In [30], an online REC-EMS considers the REC’s current condition without the future availability of generation or consumption. The majority of WEC-related research has some limitations. For instance, most studies do not consider the simultaneous forecast of energy consumption, wave peak height, time period, and direction to estimate WEC power output accurately. Additionally, they ignore the experimental testing and do not define the optimization models to enable the dual-use in grid-connected and islanded operation modes.

The contributions of this study are described as following:

• The implementation of the long short-term memory neural network (LSTM-NN) for energy forecasting, surpassing traditional persistence forecasting methods by capturing intricate temporal dependencies in the data, thus elevating energy prediction accuracy and reliability.
• Modification of a fully submerged three-tether buoy WEC hydrodynamic model, enabling real-time power output prediction using forecasted wave peak height, wave peak time period, and wave direction.
• Development of an intelligent REC EMS service that concurrently forecasts energy consumption, wave parameters, and WEC power output through a real-data-driven approach, while showcasing system versatility by functioning effectively even without an energy storage system (ESS), relying solely on super capacitor (SC) and WEC, illustrating its adaptability and resilience.

This paper is structured as follows: Section 2 discusses weather forecasting techniques and evaluates the proposed LSTM-NN with real-time wave data. Section 3 presents a modified WEC model for power output estimation using forecasted wave data. An intelligent EMS is detailed in Section 4 for power flow control. Section 5 covers a case study on the proposed EMS, and Section 6 showcases results. Conclusions and recommendations appear in Section 7.

2. Long Short-Term Memory Neural Network

The ocean waves and energy consumption data are always stochastic and dynamic by nature due to the daily fluctuation of the weather conditions and power consumption. Deterministic model-based strategies proved to be ineffective for the prediction of such datasets [31,32], since the time-series data to be forecasted are influenced by several factors, with the most significant factor being the forecasting horizon. The forecasting horizon is defined as the length of time during which output data are predicted in the future. There are four primary kinds of forecasting horizons: instantaneous forecasting (ranges from a few seconds to a few minutes), short-term forecasting (ranges from hours to days), medium-term forecasting (ranges from a few weeks to many months), and long-term forecasting (ranges from one to two years).

Depending on the time horizon of the predicted data and the computational complexity, different prediction algorithms are found in the literature [14,15,16,17,18,33,34]; among them, recurrent neural network (R-NN), multilayer neural network (ML-NN), radial basis neural networks (RB-NN), autoregressive moving average model (ARMAM), autoregressive integrated model (ARIM), exponential weighted moving average (EWMA), and LSTM-NN are the main techniques. Further classifications for LSTM-NN include (a) LSTM-NN with univariate input data and (b) LSTM-NN with multivariate input data.

The neural network (NN) known as R-NN is developed specifically to process, learn from, and forecast sequence data. With an R-NN, the output of the network from one time step is used as an input in the subsequent time step, which enables the predictions based on both direct knowledge of the previous time step and input for the current time step. The most successful R-NN algorithm is the LSTM-NN, able to create reliable models if the challenges of training a recurrent neural network R-NN will be overcome. In fact, by using both the recurrent connection of the outputs from the previous time step and an internal memory that serves as a local variable, an LSTM-NN is able to gather state across the input sequence. Unlike other models that demand that lag observations be provided as input characteristics, LSTM-NN, as a subclass of recurrent networks, can accept sequence data as an input. Figure 2 shows a detailed schematic diagram showing detailed steps for generation and load data collection, data processing, time horizon, model learning/training, and model evaluation. Algorithm used for prediction of waves parameters under study is explained in Algorithm 1.

Algorithm 1 Long short-term memory (LSTM).

Require: Time-Series Data

Ensure: RMSE of the forecasted data

1: $X\leftarrow$ 70% of Data 2: $Y\leftarrow$ 30% of Data 3: Initialize model as a sequential neural network 4: Add an LSTM layer with neurons units to model 5: Compile the model with loss function RMSE and optimizer adam 6: for each i in range(epoch) do 7:     Fit the model on X and Y for 1 epoch without shuffling 8:     Reset the states of the model 9: end for 10: return model 11: Procedure forecast_lstm(model, X) 12: $\widehat{y}\leftarrow$model.predict(X) 13: lstm_model.predict(X) 14: for each i in range(length(test)) do 15:     $X\leftarrow \mathrm{test}\left[i\right]$ 16:     $\widehat{y}\leftarrow$ forecast_lstm(lstm_model, X) 17:     Append $\widehat{y}$ to predictions 18:     expected $\leftarrow \mathrm{test}\left[i\right]$ 19: end for 20: $MSE\leftarrow$ mean_square_error(expected, predictions) 21: return $RMSE\leftarrow \sqrt{MSE}$

2.1. Model Training and Validation of Waves Data

For wave prediction, we utilized a dataset spanning from January 2004 to December 2005, which includes wave height, period, and direction. We sampled data every three hours to build time-series datasets, with 30% reserved for testing and 70% allocated for training.

Table 1. Parameters details used in LSTM model (waves and load).

Parameter	Value (Waves)	Value (Load)
Observation	710 Days [n]	60 Days [n]
Train Set	70 [%]	70 [%]
Test Set	30 [%]	30 [%]
LSTM Neurons	128 [n]	128 [n]
Batch Size	32 [n]	64 [n]
Epochs	70 [n]	150 [n]

provides details on the data and model parameters, with prediction accuracy evaluated using mean square error (MSE) and root mean square error (RMSE). Figure 3 shows the training/test data for all three datasets: wave height, wave time period, and wave direction. Also, Figure 3b shows a comparison between the expected and actual wave height where the wave heights from January 2004 to December 2005 range from 0 to 1 meters. The MSE and RMSE determined using the suggested long short-term memory (LSTM) for wave height data are 0.00725 and 0.00672, respectively. Figure 3d displays a comparison of the anticipated and actual wave time periods from January 2004 to December 2005 where the wave duration ranges from 0 to 7 seconds. Using the proposed LSTM, the MSE and RMSE for wave time period data are calculated to be 0.0159 and 0.0166, respectively. Figure 3f compares predicted and actual wave directions during the time range of January 2004 to December 2005, which varies from 0 to 360 degrees. The MSE and RMSE calculated using the suggested LSTM for wave direction data are 0.433 and 0.273, respectively. Typically, acceptable MSE and RMSE range from 0.2 to 0.5, which shows that all three time-series datasets fit within the permissible range. As shown in

Table 2. Performance comparison and analysis of the proposed LSTM.

Data Type	Persistence Forecast		LSTM Forecast
	MSE	RMSE	MSE	RMSE
Wave Height (m)	0.0607	0.0567	0.00725	0.00672
Wave Peak Time Period (s)	0.148	0.137	0.0159	0.0166
Wave Direction (°)	6.99	1.03	0.433	0.273

, the proposed LSTM are most effective at predicting wave time period when compared to wave height and wave direction.

2.2. Model Training and Validation of Energy Consumption Data

Historical energy consumption data typically follow a time-series pattern with seasonality and trends. Short-term forecasts can rely on this data, but for more accurate medium- to long-term predictions, additional factors like weather, socioeconomic, and population data play a crucial role. In our study, we use historical electricity demand data for England and Wales, complemented by weather, economic, inflation, and population data tailored to our specific needs. The electricity demand data are sourced from the National Grid ESO data library [35]. It records total metered generation, excluding specific factors, in 30 minute intervals, resulting in 48 data points per day. This dataset spans 12 years, from 1 January 2009, to 31 December 2020, with a total of 210,384 observations. We calculate the daily average demand by taking the mean of observations for each day. LSTM (long short-term memory) neural networks belong to the category of recurrent neural networks (RNN) and are frequently employed for tasks involving sequence modelling and time-series prediction. Various software libraries facilitate the implementation of LSTM networks. In our study, we utilized PyTorch 1.9, an open-source deep learning framework created by Facebook and based on Python 3.10. The dynamic computational graph of PyTorch simplifies the handling of recurrent networks such as LSTMs.

2.2.1. Many to One Time-Series-Based Energy Demand Forecasting

We employed the LSTM model described earlier to predict the next day’s electricity demand using the last 60 days of weather, population, economy, and electricity demand data. The model was configured with an “output dimension” parameter set to 1, allowing it to make single-day predictions. The training process involved using 70% of the records from the final dataset (without shuffling) for model training over 150 epochs, as detailed in Table 1. At the end of each epoch, the model predicted the next day’s electricity demand using the remaining 30% of records from the dataset (without shuffling), and the MSE was computed to assess model performance. The training and test MSE values are plotted against each epoch in Figure 4a. It is evident that both values were stabilized by the 150th epoch. The final train MSE was approximately 0.0009, while the test MSE was approximately 0.0035, indicating the model’s effectiveness in making accurate predictions. After model training, we used it to predict daily electricity demand based on 60 days of historical data. The predictions closely matched the actual data, but during the early months of 2020, there was a notable discrepancy (see Figure 4b). This deviation was due to the COVID-19 lockdown, which significantly reduced energy demand, highlighting a limitation of NN models in handling unexpected events.

2.2.2. Many to Many Time-Series-Based Energy Demand Forecasting

In this, the last 60 days of data (weather, population, economy, and electricity demand) were used to predict the next 30 days of daily electricity demand. A very similar LSTM model to the many to one time-series predictions was used to perform these forecasts, but with the “output dimension” parameter set to 30. This enabled the model to generate predictions for 30 days based on the sequence of dependent variables from the last 60 days. The training data (the first 70% of records from the final dataset, without shuffling) were used to train the model for 150 epochs. At the end of each epoch, the trained model was used to make predictions using the test data (the last 30% of records from the final dataset, without shuffling). The MSE was calculated for these predictions to understand how well the model was performing at the end of each epoch and whether the model training was improving the overall accuracy of performance. The training versus test MSE is plotted against each epoch in Figure 5a. It can be observed that both values were stabilized by the 150th epoch. The final train MSE was 0.0005, and the test MSE was 0.0092. After model training, we applied it to predict the next 30 days of electricity demand using a sliding window approach with a 60-day history. The predictions for each day were compared to the actual demand, resulting in the plot shown in Figure 5b.

3. Wave Energy Converter Power Calculation Model

3.1. Wave Spectrum and Power Density

Wave spectra are used to describe how energy is distributed at various frequencies. The pattern of the spectrum can be affected by wave height, duration, and direction. It is vital to comprehend that the spectra shown in this section are intended to explain the ocean wave spectra under extremely specific conditions, such as those that occur over a prolonged period of fixed height. A typical ocean wave will have a far more complex and varied spectrum. The ocean wave spectrum can be presented at various frequencies using Equation (1) [36].

$$S\left(f\right)=\frac{5}{16}{H}_s^2{f}_m^4{f}^{-5}\exp \left[-\frac{5}{4}{\left(\frac{f}{f_m}\right)}^{-4}\right]$$

(1)

where f is frequency. The other two parameters are the wave peak frequency $f_m$ and the significant wave height $H_s$. The power function that depends on the buoy impedance, PTO, damping coefficient, and direction of the waves can be represented by Equation (2) [36].

$$P\left(f\right)=\left[\frac{1}{4}\left[U\prime X+X\prime U\right]-\frac{1}{2}U\prime BU\right],$$

(2)

where the wave direction vector is represented by X. The damping coefficient is represented by B. Furthermore, U represents the PTO and buoy impedance-dependent velocity vector.

3.2. Power Array

To measure wave resource, a power array is a very powerful visual tool, which is an approximation of the energy in the ocean, the power obtained by a WEC, and the effectiveness of the WEC in this environment. The power array, $P_{array}$, over a range of the wave peak frequency, $f_m$, and the significant wave height, $H_s$ can be described by Equation (3) [36].

$$P_{array}(H_s,T_p)={\int }_0^{\infty }2S\left(f\right)P\left(f\right)df,$$

(3)

where $S\left(f\right)$ and $P\left(f\right)$ represent the frequency f-dependent ocean wave spectra and power function, respectively.

In Figure 6a, the ocean wave spectrum estimated with Equation (1) is shown over a range of frequencies f from 0 to 2 rad/s for varied wave heights $H_s$, with the peak time period $T_p$ set at 10 s. Figure 6a shows how the distribution of energy at various frequencies decreases when $H_s$ decreases. There is always a linear relationship between $S\left(f\right)$ and $H_s$, as evidenced by the reduction in energy distribution with changing $H_s$. Similarly, the ocean wave spectrum using Equation (1) is shown in Figure 6b for the same range of f and for varying $T_p$, while $H_s$ is taken as 0.8 m. It can be seen from Figure 6b that the distribution of energy at different frequencies is decreasing with the decrease in $T_p$. The decrease in the distribution of energy with varying $T_p$ shows that there is a nonlinear relation between $S\left(f\right)$ and $T_p$, as there is a significant shift in f. Furthermore, the power function that depends on the buoy impedance, PTO, damping coefficient, and direction of the waves calculated using Equation (2) is shown in Figure 6c. The wave direction vector is considered from 0 to 6.283 rad.

As the case study for the current scope of work, we use hourly wave data taken from MetaOcean [37]. The wave data were collected from Ko Si Chang island in Thailand (see Figure 7). For further analysis on wave behaviour, we discuss this data in quarterly changes. It can be observed that wave height in the first and fourth quarters are relatively low as compared to the wave height during the second and third quarters, as shown in Figure 7a–d. The wave height for this location usually varies from 0 to 1 m. It can be seen that the average wave peak time period in the first and fourth quarters is significantly lower than that of the other two quarters. In the second and third quarters, wave peak time period has an average value around 3 s, whereas overall data vary from 2 to 7 s. There are some instances when peak time period is either 5 or 6 s, as can be observed in Figure 7b–e. Similarly, there is a frequent variation in the wave direction from 0 to 360 degrees in the last quarter, as evident in Figure 7c–f. The wave direction varies so little between both the second and third quarters. Most of the time during these two quarters, it is around 200 degrees. If we look more closely at these two-year data, it becomes clear that there has been a movement in the yearly data, since the same wave peak heights from May to July in 2004 have moved to June to August in 2005. The same wave directional variation that occurred in September during 2004 is also moved to October in 2005 due to global climate shift and environmental changes over the years.

Power array for Ko Si Chang in Thailand is calculated using the three-tether buoy WEC model, as in Figure 1b in the Introduction Section 1 and the wave data previously discussed in Figure 7, Figure 8 and Figure 9. Calculating the power array matrix using Equation (3), we applied the $H_s$ in the range of 0 m to 0.8 m. Similarly, $T_p$ is taken from 0 and 12 s, while ${\theta }_{dir}$ is taken from 0 degrees to 360 degrees. The parameters of the simulated three-tether buoy model and the system environment parameters are presented in

Table 3. System Parameters [36].

Parameter	Value
Number of Buoy	8 n
Max. Spring PTO	55,000
Min. Spring PTO	1
Max. Damping PTO	400,000
Min. Damping PTO	50,000
Environmental Dimension	sqrt(20,000)
Water Density	1025 kg/m3
Acceleration of Gravity	9.80665 m/s2
Water Depth	30 m
Submerge Depth	3 m
Buoy Mass	3.7568 × 105 kg
Buoy Volume	523.5988 m3
Buoy Tether Angle	0.9553 rad
Buoy Radius	5 m

. With the use of the aforementioned parameters, a 3D power array visualisation is shown in Figure 8a for a range of wave height, wave peak time period, and wave direction. Similarly, a 2D power array visualization over a range of wave height and peak time period with wave direction at 180 degrees is shown in Figure 8b. It can be observed that the power output of the three-tether buoy is increasing linearly with the increase in wave height, whereas the power output visualization shows a nonlinear relationship between peak time and power output. When the peak time period is increased, the output first increases until it reaches a maximum between 7 and 10 s, at which point it begins to decrease. It shows that wave height and peak time period have a similar relationship to power output as both parameters would with energy density.

To calculate power in real time using wave data, we modified a three-tether buoy WEC model. Real-time hourly time-series data are applied to evaluate the power output using the three-tether buoy WEC model. The power output of WEC is shown in Figure 9a for the year 2004 and in Figure 9b for the year 2005. It can be observed that WEC power output varies over time with variation in wave behaviour. There are three peaks in power output for each year during the complete cycle as a result of the ideal wave height and peak time period for maximum power output occurring during those months. Comparing the power output of the two years indicates that the WEC output is also shifted, much like the wave data, as shown in Figure 7. The monthly average power output curve also highlights the shift in WEC output. We should highlight again that this may occur as a result of long-term environmental or climate changes. Note that there is a slight increase in the annual average power output in the year 2005 as compared to the year 2004, which may happen due to the increase in WEC power output during May–July in the year 2005.

4. The Energy Management Strategy

The optimization of energy resource in use is critical for REC systems; thus, an effective energy management strategy is essential. Energy management in RECs involves the optimization of different energy resources such as EMS, ESS, and generators to minimize energy costs, reduce carbon emissions, and ensure a reliable power supply. The optimization focuses on minimizing energy costs while meeting load demands seamlessly, without interruptions, and maximizing the utilization of wave energy converter-driven renewable energy sources. A comprehensive examination and comparison of the Heuristic State Flow-Based Strategy for Energy Management and the proposed optimization approach are presented in our prior work [26]. One common function of the EMS in REC is load shedding, which involves disconnecting noncritical loads during periods of high demand or low energy supply. This balances the supply and demand of energy in the REC and prevents overloading. Demand response is another strategy that incentivizes consumers to shift their energy usage to off-peak hours when energy supply is plentiful. Energy storage systems like batteries and SCs play a critical role in the EMS as they store excess energy during periods of high supply and release it during periods of low supply, balancing the load. Overall, effective EMSs are necessary for the efficient and reliable operation of REC systems. By optimizing the use of RESs, RECs can provide sustainable and resilient power to local communities.

We consider various possible operating modes based on power generation, load, voltage, current, and storage. Figure 10 shows the suggested energy management strategy with all energy units and the control group. The EMS algorithm monitors the power flows in the system by conditioning each operating mode and making decisions in accordance with the different operating modes. System-generated power, which corresponds to all power generated at the WEC farm, is referred to as $P_{SG}$. System-stored power, which corresponds to all power stored in an ESS and an SC, is referred as $P_{SS}$. Power for continuous load demand is provided by the system-generated power (SGP) and the system-stored power (SSP). Following that, the EMS is responsible for keeping the SSP group fully charged by calculating the difference between the energy generated and demanded. The EMS ensures these changes by generating reference signals for converters that immediately reflect changes. In order to maximize the amount of electrical power generated by a WEC, maximum power point tracking (MPPT) techniques in conjunction with a DC/DC boost converter are used. A DC–DC bidirectional converter is used in each of the ESS and SC groups.

The battery control system consists of the buck/boost converter as an interface for charging/discharging, control reference signal methodology, and SOC information. The battery control system stabilises the current and voltage, estimates the SOC value in real time, and holds it within the two boundaries of the minimum $SO{C}_{min}$ (for overdischarging) and maximum $SO{C}_{max}$ (for overcharging) limits. Moreover, SoC thresholds are set at 30% and 80%, respectively. The control signal to buck/boost converter is determined by calculating the difference between the WEC power and the power demanded by the load, as defined in Equation (4).

$$P_{dif f}\left(t\right)=\sum _{i=1}^N{P}_{wec}\left(t\right)-P_{load}\left(t\right),\forall t$$

(4)

where $P_{WEC}\left(t\right)$ and $P_{load}\left(t\right)$ are total power generated by N WEC and load power at time t, respectively. If the difference between $P_{WEC}\left(t\right)$ and $P_{load}\left(t\right)$ is greater than 0, the generation group will have excess power to charge the battery storage (the DC–DC bidirectional converter is in buck mode). If the difference between $P_{WEC}\left(t\right)$ and $P_{load}\left(t\right)$ is less than 0, it shows that the generation group has power deficiency, which should be taken from the storage group (the DC–DC bidirectional converter is in boost mode). The reference charge and discharge current for DC–DC bidirectional converter of the battery group for buck and boost mode is calculated as shown in Equation (5).

$$i_{b,c,d}^{*}\left(t\right)=\frac{P_{WEC}\left(t\right)-P_{load}\left(t\right)}{V_{c,d}},\forall t$$

(5)

where $i_{b,c,d}^{*}\left(t\right)$ is the reference charging and discharging current for buck/boost mode at time t, respectively. Similarly, $V_{c,d}$ is the charging and discharging voltage of the ESS group, respectively. The ESS group charge and discharge current is also limited in accordance with its manufacturer limits, as shown in Equation (8).

$$i_{max}^{*}\ge i_{b,c,d}^{*}\left(t\right)\ge i_{min}^{*},\forall t$$

(6)

where $i_{min}^{*}$ and $i_{max}^{*}$ are the minimum and maximum current limits for battery current. The control of the SC consists of buck/boost converter as an interface, control reference signal methodology unit, and voltage information. The SC control will keep the SC’s voltage value within the boundaries of the minimum voltage $V_{min}$ and maximum voltage $V_{max}$. These limits are set at 45 V and 51 V, respectively. The super capacitor group is always kept fully charged to respond rapidly to any sudden load changes. If there is no need for sudden load power for a long period, the SC group remains disconnected. When the SC group voltage is below 45 V and there is too much power in the system, the SC group bidirectional converter starts to operate in buck mode to charge the SC group. Similarly, if there is a sudden demand of the load power, the SC group bidirectional converter starts to operate in boost mode to supply the sudden load power demand. The reference charge and discharge current for DC–DC bidirectional converter of the SC group for buck/boost mode is calculated as shown in Equation (7).

$$i_{sc,c,d}^{*}\left(t\right)=\frac{P_{WEC}\left(t\right)-P_{load}\left(t\right)}{V_{c,d}},\forall t$$

(7)

where $i_{sc,c,d}^{*}\left(t\right)$ is the reference charging and discharging current for the buck/boost mode of SC group bidirectional converter at time t, respectively. The SC group charge and discharge current is also limited in accordance with its manufacturer limits, as shown in Equation (8)

$$i_{max,sc}^{*}\ge i_{sc,c,d}^{*}\left(t\right)\ge i_{min,sc}^{*},\forall t$$

(8)

where $i_{min}^{*}$ and $i_{max}^{*}$ are the minimum and maximum current limits for SC current. The primary objective of EMS strategy is to ensure the stable operation of the system by continuously monitoring the power, voltage, and current value of the storage group, power generated by WEC, and load demand power.

5. Case Study Specification

The proposed EMS algorithm (cf. Section 4) simultaneously observes WEC power generation, battery group power $P_{bat}$, SC group power $P_{uc}$, load group power $P_{load}$, SC voltage $V_{uc}$, and battery group state of charge (SOC). The EMS management of the supply and demand is based on the WEC generation, load demand, storage/SC level, and stability of the MG voltage/current output. We do consider the following case studies as shown in

Table 4. System parameters for case study.

Case	WEC Output (W)	Load Power Dammed (W)	SC F (V)	ESS V (Ahr)	SOC (%)
I	0–100	0–100	10 (48)	24 (10)	30
II	0	500	10 (48)	24 (10)	70
III	0–100	0–100	10 (48)	24 (10)	30
IV	250–500	250–500	10 (48)	24 (10)	80
V	250–500	500–250	10 (48)	24 (10)	80
VI	0–100	0–100	10 (48)	24 (10)	30
VII	500	250–500	10 (48)	24 (10)	30
VIII	250–500	500	10 (48)	24 (10)	77
IX	0–250	250	10 (48)	24 (10)	35

, where the battery’s SOC value can be low or high within the allowed range so that the SC voltage is within the allowed boundary conditions for all cases.

6. Experimental Results and Discussion

To evaluate the performance of the EMS, a real-time hardware-in-the-loop (HIL) test bench with a 24 V DC REC was utilized, as depicted in Figure 11. The optimization of load and generation forecast models employed the LSTM-NN algorithm, implemented using the Python programming language in conjunction with the Keras deep learning interface for online LSTM-NN training. Figure 12, Figure 13 and Figure 14 present comprehensive voltage, current, and power variation data across various system components. These data pertain to different load demand levels and WEC outputs, as elaborated in Section 5. For real-time monitoring purposes, a continuous information data line transmits the SC’s voltage and the battery’s SOC to the SC/battery charge controllers. In light of time constraints during experimental measurements, hourly changes in load and generation forecasts were converted to per-second changes. This adaptation enables a more effective demonstration of the EMS’s responsiveness to system fluctuations.

In Case I, as depicted in Figure 12a,b, the power output from the WEC closely matches the load power demand. The WEC output is sufficient to meet the load’s power requirements, rendering the SC and ESS inactive. Figure 12c,d represents Case II, where the WEC’s power output is nearly zero, while the load’s power demand remains at 500 W, with the ESS’s SOC at 70%. Initially, the load’s power demand is met by the ESS. However, the ESS depletes to its minimum limit of 30% at $t=14$ min, and at this point, with no available generation and the ESS also at its minimum limit, the SC starts contributing to meet the load’s power demands. Figure 12e,f depicts Case III, where both the WEC’s power production and the load demand fluctuate in the range of 0 to 100 W, and the ESS has an SOC of 30%. Initially, the load and generation are in equilibrium, rendering the ESS and SC inactive. However, as soon as generation exceeds the load demand, the surplus power is utilized to recharge the ESS since it is operating at its minimum limit. Conversely, when the load demand surpasses the generation, the ESS supplies the deficit power without activating the SC group. Figure 13a,b depict Case IV, where both the WEC’s power production and the load demand fluctuate in the range of 250 to 500 W. However, the ESS is maintained at an SOC of 80%. Initially, the load and generation are in equilibrium, rendering both the ESS and SC inactive. As soon as generation exceeds the load demand, the surplus power is used to recharge the SC, as the ESS is operating at its maximum limit. Conversely, when the load demand exceeds the generation, the surplus power is supplied from the SC until it reaches 48 V, or when there is no longer an imbalance between load and generation and the ESS has sufficient SOC to cover any difference. Figure 13c,d depict Case V, where both the WEC’s power production and the load demand fluctuate. However, the ESS maintains a stable SOC of 80%. Throughout the experiment, there are no instances where an imbalance occurs between load and generation while the ESS operates at either its maximum or minimum conditions. Consequently, SC remains inactive during the entire experiment. Figure 13e,f depict Case VI, where both the WEC’s power production and the load demand fluctuate with ESS’s SOC at 30%. Initially, the load and generation are in equilibrium, resulting in both the ESS and the SC remaining inactive. When the load demand exceeds generation, surplus power is supplied from the SC because the ESS is already at its minimum limit. Conversely, as soon as generation exceeds the load demand, surplus power is used to recharge the SC because its voltage has decreased due to the previous imbalance. Therefore, surplus power is directed to the SC until it reaches its nominal voltage of 48 V. If there is still surplus power available after this, it is then used to recharge the ESS. Figure 14a,b depict Case VII, where the load demand remains constant while the WEC’s power production fluctuates with ESS’s SOC at 30%. Initially, the load demand exceeds generation, leading to surplus power supplied from the SC because the ESS is already at its minimum limit. However, as soon as generation equals the load demand, the SC stops supplying power and becomes inactive, awaiting the next imbalance. Figure 14c,d depict Case VIII, where power generation remains constant while the load demand fluctuates with ESS’s SOC at 77%. Initially, there is no load demand, resulting in surplus power directed to the ESS as its SOC is below the maximum limit. As soon as the ESS SOC reaches its maximum limit of 80%, the surplus power is directed to charge the SC, causing its voltage to increase. However, as soon as there is load demand, all the power from the WEC is directed to meet the load demand, and stops supplying power to the SC. Figure 14e,f depict Case IX, where the load demand remains constant, but the WEC’s power production fluctuates with ESS’s SOC at 35%. Initially, the load demand exceeds generation, resulting in surplus power supplied from the ESS until its SOC reaches the minimum limit of 30%. Once the SOC reaches the minimum limit and there is still a power imbalance, the deficient power is supplied from the SC. However, as soon as generation equals the load demand, the SC stops supplying power and becomes inactive, awaiting the next imbalance.

7. Conclusions

We proposed an EMS strategy for a wave-powered renewable energy community, which is able to predict WEC generation and load demand simultaneously using LSTM-NN algorithms with high prediction accuracy, and maintain the voltage and current stability using storage groups (batteries and SCs). This approach leads to significant enhancements in mean square error (MSE) for critical variables such as wave height, time period, and direction, improving predictive accuracy by factors of 8.37, 9.30, and 16.14, respectively. The experimental results under different REC operational modes proved that the proposed LSTM-NN algorithm can improve prediction accuracy, using the same prediction dataset for WEC power output estimation. The online scheme adjusts forecast data dynamically based on current measurements, and the real-time power output of a fully submerged three-tether WEC buoy is evaluated using hourly real-time wave data. The proposed EMS architecture allows for the interaction of measurement, forecasting, and scheduling modules, with the aim of minimizing load disconnections. In the presented case studies, we showed an intelligent renewable energy community (REC) EMS service that concurrently forecasts energy consumption and WEC power output through a real-data-driven approach, while showcasing system versatility by functioning effectively even without an ESS, relying solely on an SC and a WEC, illustrating its adaptability and resilience.

8. Future Work

This study paves the way for intriguing future investigations, wherein, in future work, we can incorporate greater realism into the wave energy converter (WEC) model by considering factors such as submerged depth, buoy radius, and tethered angle. Enhanced deep neural networks can be employed to predict WEC output in specific locations, thereby assisting in system planning and operation. Despite being in the early stages of development, WEC technology is the subject of ongoing research aimed at enhancing its efficiency and cost-effectiveness. The potential for wave energy to emerge as a substantial source of renewable energy remains high, especially in coastal regions with significant wave energy potential.