Smart Dashboard for Sustainable Management of Electrical Energy in a Rankine–Hirn Power Station

Abstract

This paper highlights the eco-efficiency of a sustainable digital solution to support decision-making in resolving the problem of sudden production drops and associated energy waste in industrial power plants, especially those operating with a steam turbomachine. The solution involves a multi-interface digital dashboard that generates insightful visual reports and gives proactive alerting to the decision-makers about potential underperformances to ensure resource optimization. For the studied use case, it involves the development of three interfaces of the dashboard, so as to perform the sustainable monitoring of a thermoelectric power plant based on the Rankine–Hirn cycle as follows: the first interface is about real-time monitoring of thirty-two key physical parameters equipped with a notification system. The second interface displays the historical trends of all the plant variables, in order to help in detecting incipient abnormal deviations before they impact environmental efficiency. Lastly, the third platform covers a predictive model using the XGBoost algorithmic method to forecast the future behavior of the electrical power as the target variable of the power plant. The XGBoost method was selected after a comparative assessment which also included the algorithms of Random Forest Regressor (RFR) and Gated Recurrent Unit (GRU). As a final step, this solution was later tested in a simulation environment built under the “Node-Red” platform, through an industrial decision scenario. The concrete findings validate the framework’s sustainability metrics, demonstrating the ability of the solution to help in preserving, for each production cycle of two years, up to 7.6 GWh of electrical energy that would otherwise be wasted, which translates into a potential cost-saving exceeding 633,247.9 USD, as well as an ecological impact by preventing the emission of 4628 tons of CO₂.

1. Introduction

In a global context marked by high electricity consumption and volatile energy costs, energy efficiency has become a strategic imperative for power generation plants. In this regard, optimizing production performance is no longer an optional choice, but rather an essential issue for achieving industrial efficiency.

Within this framework, digital transformation is positioned as an unavoidable lever for achieving operational excellence across energy production facilities [1]. In fact, Industry 4.0 is fundamentally tied to smart management and real-time optimization, particularly when powered by artificial intelligence solutions.

This research work focuses specifically on the real case of a steam thermal power plant based in Morocco, which operates using the Rankine–Hirn thermodynamic cycle, where we will seek to find an approach to a recurring problem affecting this production unit.

Indeed, the central problem around which this article is structured concerns the phenomenon of sudden drops in electrical energy production within a Rankine–Hirn steam thermal power plant. These drops are often caused by the abnormal deviation of a key physical process parameter, which generates undesirable losses and impacts the performance of the plant.

The novelty of this research work focuses on the case of a steam thermal power plant, which operates using the Rankine–Hirn thermodynamic cycle, where we suggested a Decision Support System (DSS) for the recurring problem of sudden power drops affecting this kind of production unit. In fact, we have developed a predictive model through “Python 3.11” and the historical archived data of the power plant, then we associated it to a visual real-time dashboard built under “Node-Red” in order to alert proactively the end-users about this kind of underperformance and help them to avoid power loss guaranteeing sustainable monitoring of the plant.

As a solution to this underperformance, we have chosen Industry 4.0 as a strategic axis to propose a modern decision support system (DSS). To address this challenge, we will concentrate on establishing an end-to-end digital solution, based on the valuation of the historical data archived in industrial IT servers [2]. It aims to help the leaders of this kind of energy infrastructure in their decision-making process regarding the phenomenon of unpredictable production drops. Consequently, this solution should guide the end-users of the power plant to act proactively on the root causes of these deviations, thereby preventing their negative impact on productivity.

In order to effectively reach this objective of resolving the critical issue of sudden power drop in Rankine–Hirn power plant, we have determined that the structure of the present work will follow a logical progression according to the following sections:

Section 2 will be devoted to exploring the major research works related to our scientific domain, namely business analytics in the electric power generation sector. Each work will be characterized by the category of the studied power plant, its strengths, and its limitations. Thus, positioning our topic relative to peers in the field will ensure our work is relevant and establishes the necessary framework to deliver the expected solution.

Section 3 will focus on a scientific framing of our application case linked to the Rankine–Hirn power plant, defining the process steps and the energy assets used for producing electricity from high-pressure steam. Similarly, we will present the methods used to develop the targeted decision support system (DSS), which will be materialized by a multi-interface digital dashboard, representing the core of our work’s solution.

Furthermore, knowing that the main objective of the solution is to predict undesirable power drops, one of the interfaces will be supposed to show the results of a predictive model based on machine learning. Consequently, this section will be an opportunity to define the physical input variables for modeling, so as to forecast the electrical power as an output variable, as well as the algorithmic method chosen to elaborate the model.

Section 4 and Section 5 will describe the results obtained, showing the dashboard interfaces that will ensure a smooth user experience with an intuitive working environment, both in terms of ergonomics and simplicity of navigation. This will accelerate access to the main information, helping in establishing fast and correct decision-making, in order to enable proactive actions to be taken before the problem of power drops occurs.

Similarly, we will opt for the creation of a mobile version of the dashboard to take advantage of the benefits afforded by Industry 4.0 technologies like IoT devices, particularly in terms of remote and real-time access to industrial information, which further promotes the decision-making improvement process. Likewise, this part will provide an opportunity to test and experiment the use of the dashboard in a simulation environment, so as to quantify its potential gains in terms of energy optimization, cost-saving, and environmental protection.

2. Recent Works on Industry 4.0 Methods for Improving Power Generation Efficiency

Before delving into the heart of the subject, it is important to explore the recent research works related to this field. Indeed, the objective is to provide an overview of the current state linked to the scientific and technological literature about decision support systems, which would be adapted to the context of Industry 4.0, and designed to serve thermoelectric power plants using turbomachinery.

As indicated in

Table 1. Summary of recent related works.

Author	Year	Category of Power Production	Main Contributions and Impacts	Ref.
Harmoko	2026	Steam Turbine Power Plant	The interactive dashboard of this paper features a predictive algorithm for the power output of a turbine generator. This model is built using the K-Nearest Neighbor (KNN) algorithm. It was an effective decision-making tool, distinguished by the fact that it offers up to 8% additional productivity simply by using the dashboard to make proactive decisions based on the model’s predictive results. Our research addresses a more specific and critical operational challenge. Unlike existing general models, this work focuses on the Rankine–Hirn thermodynamic cycle within steam thermal power plants. We introduce a Decision Support System (DSS) specifically engineered to tackle sudden power drops. By integrating a custom Python-based predictive model with a real-time Node-RED interface, our approach moves beyond general productivity to provide a robust, sustainable monitoring solution that alerts users to specific underperformance events before they occur. While the KNN dashboard offers a valuable 8% productivity gain, this work elevates the domain by shifting from broad optimization to targeted risk mitigation within the specialized Rankine–Hirn cycle. By replacing a standalone tool with an integrated, real-time Decision Support System, it bridges the gap between passive predictive modeling and proactive industrial maintenance.	[3]
Rediske	2025	PV Solar Power Plant	In this case, the paper has demonstrated the value of automated dashboards for solar power parks, primarily focusing on the calculation of Key Performance Indicators (KPIs) to improve asset availability and streamline maintenance. However, while KPI monitoring is essential for assessing historical performance, it often lacks the predictive foresight required to manage complex thermodynamic fluctuations. Our research advances this field by transitioning from passive KPI tracking to a proactive decision-making system tailored for steam thermal power plants. Unlike the aforementioned solar-focused studies, our approach utilizes Python-driven predictive modeling to specifically address the problem of sudden power drops in the Rankine–Hirn cycle. By integrating this model into a real-time Node-RED dashboard, we provide a solution that does not just report failures but actively alerts operators to prevent imminent power loss. While solar dashboards effectively track historical KPIs for asset availability, they lack the predictive foresight needed for complex thermodynamic systems. This study successfully bridges that gap by shifting from passive reporting to active, real-time failure prevention within the Rankine–Hirn cycle.	[4]
Bawane	2025	Wind Power Plant	It is an IoT-based dashboard to realize real-time management of wind farms. This mode of monitoring has allowed us to ensure a high efficiency of power production, knowing that it also includes an interface of power prediction based on machine learning algorithms that makes it possible to increase daily productivity by more than 3 MWh. Our study diverges from these general IoT monitoring systems by focusing specifically on the Rankine–Hirn thermodynamic cycle. Rather than focusing solely on incremental productivity gains, it lies in the development of a predictive model designed to mitigate sudden power drops—a recurring issue in steam plants. By leveraging historical archived data, we provide a specialized solution for real-time alerting and long-term operational sustainability that generic IoT frameworks do not address.	[5]
Hamza Muhammad	2024	PV Solar Power Plant	This research leverages the transformative potential of Industry 4.0 by integrating emerging technologies, such as the MQTT (Message Queuing Telemetry Transport) protocol and high-precision IoT sensors, to construct a comprehensive real-time dashboard. The primary objective is to streamline the decision-making process by providing stakeholders with instantaneous visibility into operational metrics, thereby significantly enhancing time management and resource allocation. By digitizing the feedback loop, the system reduces the latency between data acquisition and corrective action, which is vital in fast-paced industrial environments.	[6]
Tohir	2024	Coal-Fired Power Plant	This work presents a low-investment, high-impact framework that capitalizes on the deployment of smart sensing technologies, specifically intelligent energy meters, to establish a centralized and remotely accessible monitoring dashboard. By prioritizing cost-effective hardware integration, the solution demonstrates how legacy industrial infrastructure can be modernized without the need for prohibitive capital expenditure. This study was instrumental in the ISO 50001 certification process of a coal-fired power plant, where the implementation of these enhanced decision-making tools directly contributed to a documented 11.76% reduction in energy losses. This proves that real-time visibility is a powerful catalyst for operational efficiency and rigorous energy management. This low-cost hardware framework effectively reduces energy losses by maximizing real-time visibility of current waste. Our study builds upon this foundation by integrating predictive modeling, moving beyond mere monitoring to actively prevent future operational drops.	[7]
Arief	2024	Wind Power	The particularity of this research work is that it is based on the “Node-Red” and Remote-Red platform, to design a mobile dashboard that can be consulted from smartphones. This architectural choice facilitates a transition toward “On-the-Go” industrial management, allowing engineers to bypass the constraints of fixed-site monitoring stations. By centralizing data through this agile, flow-based programming environment, this idea guarantees rigorous monitoring of the operating parameters of the wind farm studied, which saves time and efficiency by reducing decision-making times for managers. Such a system proves invaluable for remote sites, where the ability to visualize real-time performance metrics on a handheld device can mean the difference between a minor adjustment and a costly system failure. While mobile dashboards successfully liberate operators from fixed workstations through real-time visualization, they remain focused on passive data access. This study advances that mobility by embedding predictive alerts directly into the handheld interface, transforming a remote viewing tool into a proactive decision-making asset.	[8]
Cigánek	2024	Steam Turbine Power Plant	In the case of a steam turbine and using the “Node-Red” platform, the study demonstrates that the combination of Industry 4.0 technologies and advanced automation can increase the efficiency of operational monitoring. By integrating real-time data flows with intuitive visual programming, this approach allows for a granular oversight of thermodynamic variables that were previously siloed in legacy systems. Consequently, decision-making becomes more fluid, which also has an impact on sustainability and decarbonization. This fluidity enables operators to maintain the turbine within its optimal efficiency window, thereby reducing fuel consumption and minimizing the carbon footprint associated with thermal power generation.	[9]
Al Mohannadi	2024	Gas Turbine Power Plant	This study incorporates a predictive model of power production based on Random Forest Regressor (RFR), built using a five-year historical dataset. By leveraging this ensemble learning technique, the research captures non-linear relationships between operational variables and output, ensuring a robust baseline for performance forecasting. The model’s results were then incorporated into an interactive dashboard, transforming complex algorithmic outputs into actionable visual insights for plant operators. This achievement has improved the decision-making process, both in terms of energy and the environment, by reducing greenhouse gas emissions by approximately 1,400 tons per year, primarily through the optimization of fuel combustion and the early detection of efficiency drifts. An analysis of this RFR-based model highlights its success in reducing emissions by utilizing a five-year dataset to optimize fuel combustion. Our work builds upon this insight by introducing a multi-algorithm approach and real-time integration to target sudden, short-term power drops within the specific Rankine–Hirn cycle.	[10]
Kurniawan	2024	Combined Cycle Power Plant (CCPP)	The mechanism used to develop the smart dashboard in this case was Node-RED associated with a predictive model developed under Python. This hybrid architecture leverages the rapid prototyping and data-flow orchestration of Node-RED alongside the robust analytical power of Python’s machine learning libraries. By creating a seamless bridge between raw data acquisition and advanced algorithmic processing, the system transforms static metrics into dynamic forecasts. Even if the impact of this study was not clearly quantified, the developed smart dashboard allows to maximize the reliability of the plant. This is achieved by shifting the operational paradigm from reactive maintenance to a more informed, proactive stance, ensuring that critical components are monitored against their theoretical performance curves. This became true thanks to the intelligent and instantaneous notifications that enormously helped the site managers to make good decisions at the right moment, effectively reducing the cognitive load on operators during high-pressure scenarios.	[11]
Khan Malhar	2024	PV Solar Power Plant	The target behind developing the business intelligence dashboard of this paper was to enhance the maximum power point tracking (MPPT) to achieve an efficient solar power generation. By providing a centralized platform for monitoring atmospheric variables such as irradiance and ambient temperature, the dashboard serves as a strategic interface to optimize the alignment between photovoltaic output and load demands. The dashboard was equipped with an artificial intelligence model based on artificial neural networks (ANN) that has given an accuracy score of 99.2%, which was a tangible added value to the dashboard as a decision support system. This high-precision forecasting allows operators to anticipate fluctuations in energy harvest with near-certainty, transforming the dashboard from a simple visualization tool into a robust engine for grid stability and performance optimization.	[12]
Mazdak	2024	Steam Turbine Power Plant	Increasing the overall efficiency of the power plant was the main objective of this work. To achieve this, the research focused on a sophisticated technological framework performed through a solution involving supervision and monitoring aiming to maximize energy recovery from the gas at the chimney output. By deploying advanced thermal sensors and real-time data acquisition systems, the study targeted the “waste heat” typically lost to the atmosphere, seeking to reintegrate this energy into the plant’s thermodynamic cycle. The implementation of this solution could contribute to enhancing the overall efficiency from 83% to 91%, representing a massive leap in operational performance and fuel economy. This gain demonstrates the untapped potential of digital oversight when applied to the recovery of low-grade thermal energy from flue gases.	[13]
Mohapatra	2024	Diesel Power Generation	The authors of this work have chosen the technology of the Internet of Things (IoT) to build their decision-making solution based on business intelligence. By deploying a network of interconnected sensors across the plant’s critical infrastructure, they have established a continuous data pipeline that feeds a centralized analytical engine. In this sense, the dashboard was equipped with a model for predictive maintenance, allowing breakdowns to be anticipated by predicting the remaining useful life (RUL) of key components such as turbines, pumps, or bearings. This transition from traditional scheduled maintenance to a condition-based approach allows the user to act proactively to gain energy productivity, ensuring that assets are repaired only when necessary but before a catastrophic failure occurs, thus minimizing unplanned downtime and optimizing the levelized cost of energy. An analysis of this IoT-driven framework reveals how predicting the remaining useful life (RUL) of components successfully minimizes unplanned downtime through condition-based maintenance. Our work builds upon this predictive approach by shifting the focus from individual component wear to the real-time prevention of sudden thermodynamic power drops across the entire Rankine–Hirn cycle.	[14]
Zongjie	2024	Steam Turbine Power Plant	The BI platform used in this work is centered around a predictive model of electrical energy production, driven by the long short-term memory (LSTM) algorithm. By utilizing this specific type of recurrent neural network, the system is uniquely capable of capturing long-term dependencies and complex temporal patterns within the power plant’s historical data, which traditional linear models often overlook. In addition, this work is particularly notable for the addition of an alarm system, alerting users to abnormal deviations in process variables. This proactive layer ensures that any drift from the optimal operating envelope is flagged instantly. This approach aims to improve end-user responsiveness to undesirable situations that could impact productivity, which in itself represents an effective decision-making tool that allows productivity to be increased three times compared to standard monitoring, transforming the dashboard from a passive display into an active guardian of operational continuity. Examining this LSTM platform reveals its strength in capturing temporal patterns and boosting responsiveness through basic process alarms. Our research expands on this concept by merging predictive intelligence with a three-tiered alert system to actively mitigate sudden thermodynamic power drops.	[15]
Thirumurthy	2024	Gas Turbine Power Plant	This paper highlights the increased value of gas turbines through digitalization and real-time data analysis, emphasizing how silicon-based intelligence now complements traditional steel-based engineering. It is founded on the following three digital applications co-developed by Siemens Energy: First, remote diagnostics to increase availability and reduce costs by enabling offshore or isolated sites to access global expert centers without the need for physical travel. Second, advanced vibration analysis for reliability-driven maintenance, as well as maintenance optimization and reduced lifecycle costs. Third, by utilizing high-frequency telemetry to detect spectral anomalies in rotating components, the system can identify incipient flaws long before they lead to mechanical failure. The study demonstrates how these tools are transforming asset management by enabling proactive decision-making and operational efficiency for customers, effectively shifting the industry standard from reactive “break–fix” models to data-driven strategic planning.	[16]

, the papers are organized by year of publication, author, the energy production category, and the major contributions of each work.

An article-by-article analysis was conducted to ultimately determine the most appropriate algorithmic methods for the predictive model of our specific case study.

In light of this analysis, we will henceforth focus on proposing a data-based system of decision-making, in relation with a power plant based on the Rankine–Hirn thermodynamical cycle. This specific case needs to be more developed by researchers, which justifies our research direction.

Table 1 provides a comprehensive benchmark of recent algorithmic applications across comparable thermal energy systems. While these studies demonstrate the efficacy of machine learning (ML) within specific domains, a distinct technological and scientific gap remains: a rigorous, direct comparative analysis of these advanced architectures has not yet been established for high-pressure turbogenerator systems subjected to highly fluctuating operational variables.

To address this gap, this study leverages the benchmarking insights from Table 1 to evaluate three of the most prominent predictive methods identified in high-pressure turbine applications (such as CSP, biomass, and Brayton-cycle gas stations): Random Forest Regressor (RFR), Extreme Gradient Boosting (XGBoost), and Gated Recurrent Unit (GRU). By systematically comparing a tree-based ensemble, a gradient-boosted framework, and a recurrent neural network under identical operational constraints, this work establishes a definitive technical foundation to determine the optimal architecture for predicting turbogenerator power degradation.

Similarly, all recent research in this field emphasizes the necessity of timely decision-making without delay. Consequently, it has become essential to associate predictive models with dynamic and real-time dashboards. To this end, “Node-Red” has emerged as a leading solution, proving its effectiveness as a validated simulation environment.

In the following section, we will present a scientific framework for the specific case of the power plant under study, as well as the architecture that will allow us to link the predictive model with the dashboard designed in “Node-Red”.

3. Materials and Methods

3.1. Scientific Framework of the Case Study

The process of power production involved in this study is a closed water-steam conversion cycle. As shown in Figure 1, it is a cycle composed of five thermodynamic transformations; it concerns the improved Rankine cycle, also called Rankine–Hirn [17].

In order to understand the physical principle of the energy exchange at each of the five stages, we apply the first law of thermodynamics as follows [18]:

∆E_sys = ∆U_int + ∆E_pot + ∆E_kin = W_k + Q_h

(1)

where

• ∆E_sys: variation in the microscopic and macroscopic energy of the system;
• ∆E_pot: variation in the potential energy of the system;
• ∆E_kin: variation in the kinetic energy of the system;
• ∆U_int: variation in the internal microscopic energy;
• Q_h: heat transfer exchanged between the system and its surroundings;
• W_k: mechanical work exchanged between the system and its surroundings.

However, the framework of our study highlights the special case where the closed system is assumed to be at rest in a Galilean frame of reference. It is then a motionless system, where the speed and the height of its equipment do not change during the transformation processes [19]. Consequently, we deduce the following physical equation:

∆U_int = W_k + Q_h

(2)

By introducing the physical notions of the mass flow (ṁ) and the specific enthalpy (h), we could express the power exchanged in kilowatts for each transformation by the following equation:

ẇ_ij + q^·_ij = ṁ_ij × ∆h_ij

(3)

where

• ẇ_ij: the power of the work exchanged between the two states “i” and “j” (kW);
• q^·_ij: the power of the heat exchanged between the two states “i” and “j” (kW);
• ṁ_ij: the fluid mass flow rate of the between the two states “i” and “j” (kg/s);
• ∆h_ij: the difference between specific enthalpies in states “i” and “j” (kJ/kg).

Consequently, in

Table 2. Physical equations of the power exchanged in Rankine–Hirn cycle.

Steps of the Process	Power Machine	Particularity of the Transformation	Physical Equations of Power Exchange
From state (A) to state (B)	Water pump	Adiabatic	ẇAB = ṁAB × (hB − hA)	(4)
From state (B) to state (C)	Steam boiler	Isobaric	q·BC = ṁBC × (hC − hB)	(5)
From state (C) to state (D)	Steam superheaters	Isobaric	q·CD = ṁCD × (hD − hC)	(6)
From state (D) to state (E)	Turbine	Adiabatic	ẇDE = ṁDE × (hE − hD)	(7)
From state (E) to state (A)	Steam condenser	Isobaric	q·EA = ṁEA × (hA − hE)	(8)

, we obtain the equation related to each of the five equations of every thermodynamical transformation [20].

The calculation of the power exchanged during the five transformations of the Rankine–Hirn cycle is important for many reasons. First, it allows us to determine the network and to know the useful mechanical energy produced by the cycle. Secondly, it gives the possibility to calculate the thermal efficiency of the cycle and to measure the efficiency of conversion of heat into work. Last but not least, it allows us to size the energetic machines by providing the necessary nominal data for the design-like power and geometric sizes [21].

3.2. Architecture of the Designed Dashboard

The data source used to feed the dashboard by process information is an OPC virtual server called “OPC Prosys 5.6.0”, which simulates the same principle of a real OPC server. It allows us to produce the power plant data in real time [22].

As indicated in the architecture of Figure 2, “Node-Red 4.1.9” is the software tool that we have adopted to build this dashboard. It contains the following two environments: “Back-end” and “Front-end” [23].

In addition, the flow of nodes includes a connection between the dashboard and the machine learning model running on a software platform supporting Python language like “Visual Studio Code 1.87.1”. This connection is made by means of a node called “Python-Shell”, which receives the simulated data from the same data source of “OPC Prosys” [24].

Additionally, the use of the software “Symbol Factory” affords to introduce professional graphic symbols to the dashboard, especially those related to the industrial machines and exchangers, like boiler, turbine, and electrical generator [25].

On the far right of the architecture, we have inserted a block dedicated to the instant notifications system. The purpose of this block is to boost the interaction of the dashboard with its end-users by sending them alerts related to the limit thresholds of the power plant parameters using three different ways of communication as follows: the appearance of local alarms on the interface of the dashboard, sending notifications by email, as well as sending them through SMS messages using the “Push-Bullet 18.12.1” software [26].

The last block concerns the remote access to the dashboard by using a mobile application attached to “Node-Red” called “Remote-Red”. The said application is connected to the dashboard by scanning a QR code generated by a node called “Remote-access” [27]. This solution enhances the attractivity of using IoT devices like smartphones to benefit from their advantages, like real-time reporting and mobility [28].

Regarding the gray block related to the node flow created under “Node-Red”, we described it in detail in Figure 3. Indeed, it consists of six sub-blocks that we have surrounded by different colors as follows:

The nodes surrounded by blue represent the physical variables of the Rankine–Hirn thermoelectric power plant. Their role is the real-time data injection from the OPC UA virtual server, while the node surrounded by orange constitutes the main client database to feed the dashboard with real-time data.

The nodes encircled in green are devoted to displaying data in real time on the dashboard, while the nodes encircled in red are dedicated to setting instantaneous notifications and alarms in case of deviations. Similarly, to ensure remote access to the real-time parameter variations, we have used the node surrounded by gray called “remote-access”.

Finally, to improve the graphical appearance of the dashboard, we used the nodes encircled in purple to display the clock, the banner, and the background.

3.3. Developed Predictive Model of the Third Interface

This section concerns explaining the content of the third interface of the interactive dashboard, which is devoted to showing the results of the predictive model based on AI.

In fact, these results are provided continuously by a machine learning model, which is built under the Python programming language. As shown in

Table 3. Nominal values and alarm thresholds of the dataset features.

Input Variables	Symbol	Physical Unit	Nominal Value	Alarm Value
Pressure of the fluid at state (A).	PA	bar	1.05	1.2
Temperature of the fluid at state (A).	TA	°C	77.5	90
Pressure of the fluid at state (B).	PB	bar	66	71
Temperature of the fluid at state (B).	TB	°C	108	120
Pressure of the fluid at state (C).	PC	bar	60	68
Temperature of the fluid at state (C).	TC	°C	277	290
Pressure of the fluid at state (D).	PD	bar	62	68
Temperature of the fluid at state (D).	TD	°C	397	410
Pressure of the fluid at state (E).	PE	bar	1.05	1.2
Temperature of the fluid at state (E).	TE	°C	98.2	110
Mass flow of the fluid during the transformation (A→B).	ṁAB	t/h	210	220
Rotation speed of the pump.	SAB	rpm	1480.1	1450
Electrical current of the motor.	IAB	Ampere	86.5	82
Level of the water tank.	LAB	%	75.0	15
Mass flow of the fluid during the transformation (B→C).	ṁBC	t/h	148	160
Input temperature of the heating gas at the boiler.	TBC1	°C	448	465
Output temperature of the heating gas at the boiler.	TBC2	°C	298	315
Mass flow of the fluid during the transformation from (C→D).	ṁCD	t/h	161	170
Input temperature of the heating gas at superheater n°1.	TCD1	°C	346	360
Output temperature of the heating gas at superheater n°1.	TCD2	°C	222	235
Input temperature of the heating gas at superheater n°2.	TCD3	°C	238	255
Output temperature of the heating gas at superheater n°2.	TCD4	°C	174	185
Mass flow of the fluid during the transformation from (D→E).	ṁDE	t/h	223	235
Rotation speed of the turbomachine.	SDE	rpm	3000	2850
Mass flow of the fluid during the transformation from (E→A).	ṁEA	t/h	126	135
Temperature of the cooling water at the condenser.	TEA	°C	18	20
Pressure of the vacuum at the condenser.	PEA	bar	−0.85	−0.75
Level of the cooling water basin.	LEA	%	75.0	15
Electrical voltage of the power generator.	UALT	kV	10.0	10.5
Electrical current of the power generator.	IALT	Ampere	2885	3050
Power factor.	FP	without unit	0.85	0.79
Frequency of the electrical current.	F	Hz	50.0	50.5

, this model is to be animated through the historical data of thirty-two input physical variables (features) and one output variable (target), which is the electrical power produced by the generator.

Before animating the model, we performed a two-step preprocessing of the two-year data history to increase the quality and reliability of the used data as follows:

The first step involved eliminating outliers such as missing values, negative values, non-numeric values, and values recorded during production shutdowns.

The second step focused on the correlation study, to check whether all input variables (features) have a strong dependent relationship with the output variable (target).

After that, the definitive prepared dataset was used to animate the model using the three different algorithms that are: RFR, XGBoost and GRU. To perform this, we divided the dataset into the following two parts: the first, which consists of 80% of the data starting from 1 January 2024 to 7 August 2025, is intended for the model’s training phase; and the second, which consists of 20% of the data starting from 8 August 2025 to 31 December 2025, was used to perform the testing phase to confirm the model’s reliability, knowing that the data sampling period was equal to 1 s [29].

Following that, it was time to evaluate the performance of the developed model. To achieve that, we used three statistical metrics to measure this performance, which are the score of determination which should be close than 100%, the mean square error, and also the mean absolute error. For these two errors, the closer they are to zero, the better it is for the model. To put it simply, we have summarized the different results of assessment in

Table 4. Results of evaluation metrics using RFR, XGBoost and GRU algorithms.

Metrics of Assessment	Mathematical Formula		RFR		XGBoost		GRU
			Training Phase	Testing Phase	Training Phase	Testing Phase	Training Phase	Testing Phase
Score of determination	${\mathrm{R}}^2=1-{\displaystyle \frac{{\displaystyle {\sum }_{k=1}^n}{(P_k-\widehat{P_k})}^2}{{\displaystyle {\sum }_{k=1}^n}{(P_k-{\overline{P}}_k)}^2}}$	(9)	0.9204	0.9304	0.9936	0.9911	0.9633	0.9721
Mean Square Error	$\mathrm{M}\mathrm{S}\mathrm{E}={\displaystyle \frac{1}{N}}{\displaystyle {\sum }_{k=1}^n}{(P_k-{\widehat{P}}_k)}^2$	(10)	0.0341	0.0837	0.00024	0.0003	0.0065	0.00298
Mean Absolute Error	$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{N}}{\displaystyle {\sum }_{k=1}^n}{|P}_k-{\widehat{P}}_k|$	(11)	0.0793	0.0776	0.0031	0.00435	0.00783	0.00801

, also including the mathematical equations of each used metric [30].

The significations of each parameter of the three equations are:

• $P_k$: the measured value of the output variable P_ALT;
• $\widehat{P_k}$: the predicted value of the output variable P_ALT;
• ${\overline{P}}_k$: the average value of the output variable P_ALT;
• N: the number of data points of the set (training set or testing set).

As we can observe, the evaluation results obtained after completing the model’s training and testing phases demonstrate a very good level of performance, allowing the developed model to forecast the electrical power P_ALT produced by the turbomachine in a highly reliable manner.

It is noted that the highest performance was achieved using the XGBoost algorithmic method, which yielded the best R² score and the lowest MSE and MAE errors compared to those generated by RFR and GRU, during both the training and testing phases.

In summary, this evaluation demonstrates that the XGBoost algorithm exhibits the best generalization capabilities on new data, without showing signs of either overfitting or underfitting. The superior performance of the XGBoost model is also attributable to the hyperparameter tuning summarized in

Table 5. Hyperparameters of the XGBoost model.

Hyperparameter	Value
Nbr. of estimators	500
Learning rate	0.02
Max_depth	7
Early stopping round	50
Subsample	0.7

. Consequently, we have selected it as the core algorithmic method for our application’s predictive model.

4. Results and Discussion

This section is more focused on transforming raw data into actionable analytics in real time. This is achieved through business intelligence (BI) techniques, especially via the interactive dashboard that we have described before.

4.1. Findings

As a major achievement for this paper, we have succeeded in developing a real-time dashboard with three interfaces for the power plant under the IT tool of “Node-Red” [31]. Among the features of this dashboard, there is its ability to automatically retrieve the industrial data in real time, making it a genuine decision-making tool readily available to decision-makers.

Additionally, to benefit from the advantages of IoT devices brought by the era of digital transformation, we were able to develop a mobile version of the dashboard under “Remote-Red”, containing the same interfaces and allowing the property of mobility to be added to the work performed, which promotes making decisions at the right time without any delay.

The effectiveness of this digital solution rests on several critical assumptions. First, it is assumed that the data integrity of the thirty-two physical sensors is maintained through regular calibration, as any sensor drift would propagate errors through the Node-RED interface. Second, the model assumes seamless network connectivity; specifically, that the Remote-RED mobile platform has access to a low-latency connection, which is not always guaranteed in isolated industrial zones. Third, regarding the quality of data, it is assumed that the historical archived data used to train the underlying models is free from significant outliers or “data gaps” caused by sensor outages, and that it accurately represents the full operational range of the Rankine–Hirn cycle. Finally, concerning the simulation environment, there is an underlying assumption that the software-defined testing conditions in Python and Node-RED accurately mirror the physical stresses and high-temperature variability of the real-world thermal plant, ensuring that the DSS remains reliable under actual industrial loads.

In Figure 4, we can observe the interface that is designed to display the real-time variations in all the physical variables of the power plant, notably the values of thirty-two input variables (features), as well as the model’s output variable (target), which is the electrical power generated by the turbomachine [32]. Similarly, to provide more visibility on the outcome of this generated power, we have also added the following values: the power consumed internally by the plant P_INT and the power exchanged with the grid P_GRID.

Similarly, the second interface, which is displayed in Figure 5, allows us to visualize the historical variation curves for all the predefined variables. It provides a clear understanding of the temporal behavior of each variable and enables a precise assessment for detecting variables that exhibit abnormal deviations which go against the production interests of the power plant.

Likewise, the third and last interface, which is represented in Figure 6, shows the prediction results generated by the AI-based model developed in Section 3, illustrating the smart aspect of the dashboard. Indeed, these results are shown through a diagram that displays the following two superimposed curves: the prediction curve and the curve of the actual measurements of the power produced P_ALT.

We also note that we have chosen two prediction horizons for the following reasons:

• The first horizon is for the short-term and concerns 24 h ahead of forecasting, it is specially intended for the variables that are actionable during the full running mode of the power plant [33].
• The second horizon is for mid-term forecasting, it involves 7 days of prediction ahead, and it is intended for the variables that are actionable during a shutdown of the power plant.

The approach followed is geared toward leveraging scheduled maintenance shutdowns, which are programmed for one day each month, to include any desired proactive actions in the list of planned tasks, completing them during hidden time (background time) during these outages.

To sum up, this approach of 7 days ahead of forecasting aims to avoid shutdowns specifically dedicated to the desired action, which would undoubtedly be penalizing.

Despite this result, the predictive model may suffer from selection bias if historical data excludes extreme operational anomalies, alongside overfitting that limits its reliability during unprecedented plant states. Uncertainties also arise from sensor noise and data drift, where physical aging of the Rankine–Hirn components causes the model’s outputs to deviate from the plant’s actual real-time performance.

Let us continue with business intelligence (BI) techniques, but this time by applying them in IoT devices, where we developed a mobile version of the dashboard, as shown in Figure 7. This version will help managers make their decisions wherever they are, ensuring responsiveness in decision-making, which leads to better performance [34].

4.2. Test and Experimentation Through Simulation

To fully evaluate the operational usefulness of the proactive actions dictated by the prediction results of the model, it is necessary to explain them through a decision scenario. To do this, let us consider the example of a prediction result, which indicates that energy production will fall below the critical value of 33 MW in 24 h. As shown in Figure 8, this value represents the nominal consumption of the various loads of the power plant, as well as the plant to which it belongs, representing the surroundings of the Rankine–Hirn cycle with which it exchanges heat and mechanical work.

Given that the turbo-alternator unit under study is connected to the public electricity grid, there are two possible cases in this sense:

-

Power plant with energy excess: when production exceeds the critical value of 33 MW, the excess of energy is injected into the grid.

-

Power plant with an energy deficit: when production falls below the critical value of 33 MW, the energy deficit is spontaneously compensated by the grid.

For the second case, which represents an undesirable event for the managers of the thermoelectrical power plant, this was the subject of the predictive result in Figure 8. In fact, the objective of the prediction is to be able to react proactively to prevent the occurrence of such an event, by controlling which input parameter of the model is experiencing abnormal variations causing this underperformance.

Regarding the chosen example of the previous figure, we examined the dashboard’s interface of “real-time” and we found that it is the “power factor” parameter that begins to vary abnormally, experiencing a gradual decrease from 0.83 to 0.79.

As a proactive action, we adjusted the excitation current of the turbomachine alternator, and we observed an increase in the power factor, which allowed us to increase the predicted electrical power output.

This situation made it possible to avoid the current undesirable drop of the produced power, referring to the ability of the predictive model to ensure energy efficiency of the Rankine–Hirn power plant.

In comparison to the existing literature, our work distinguishes itself by providing a comprehensive, physics-informed Decision Support System (DSS) that transcends simple connectivity or descriptive monitoring. While studies such as Arief (2024) [8] and Cigánek & Dávid (2024) [9] demonstrate the utility of Node-RED for data visualization and communication, they focus primarily on the infrastructure layer rather than advanced predictive diagnostics. Similarly, while Bawane et al. (2025) [5] and Khan et al. (2024) [12] utilize AI for wind and solar optimization, these models are designed for intermittent resources and do not address the high-inertia complexities of the Rankine–Hirn cycle found in steam turbomachinery.

Unlike the work of Mohapatra et al. (2024) [14], which prioritizes parameters for diesel generators, or Kurniawan et al. (2024) [11], who focus on mechanical greasing systems, our research addresses the fundamental problem of sudden production drops through a multi-tier analytical approach. By benchmarking XGBoost against RFR and GRU, we achieve a higher degree of forecasting precision than general Industry 4.0 dashboards (e.g., Tohir et al., 2024; Hamza, 2025) [6,7].

4.3. Projected Operational Impacts of the Solution

In order to quantify the potential gain offered by the digital solution presented in this article, we identified the periods during which production fell below the critical power threshold of 33 MW. In fact, knowing that each production cycle of the power plant lasts two years, we chose the historical data from the years of 2024 and 2025 to identify these production drops.

The added value of using the digital dashboard, which is powered by the predictive machine learning model, lies in the following three aspects: energetic, economic, and environmental.

4.3.1. Calculation of the Avoidable Energy Waste

It was calculated using the following equation:

$${\mathrm{E}}_{\mathrm{W}}={\displaystyle {\sum }_{k=1}^{12}}{E}_{Wk}$$

(12)

where

• E_W: the annual avoidable energy imported from the grid, which is considered as an energy waste for the power plant (MW);
• E_Wk: the monthly avoidable energy imported from the grid (MW);
• k: is a natural integer belonging to the interval [1,2,3,4,5,6,7,8,9,10,11,12] and represents the 12 months of the year.

In

Table 6. Inventory of the imported energy from the grid due to the production drop during 2024 and 2025.

		Jan	Feb	Mar	Apr	May	June	July	Aug	Sept	Oct	Nov	Dec	Total
		EW1	EW2	EW3	EW4	EW5	EW6	EW7	EW8	EW9	EW10	EW11	EW12	EW
Imported Energy from the grid (MWh)	Year of 2024	984.8	429.4	291.5	562.8	186.7	1194.2	1181.6	895.7	0	0	0	412.1	6138.8
	Year of 2025	0	0	0	0	210.9	290.0	40.3	0	0	386.9	178.6	379.4	1486.1

, we have listed the total energy imported during each month, as well as the sum of the annual energy imported from the electrical grid during the two years, which is equal to:

E_W = 7624.9 MWh

Likewise, the energy loss calculation enabled us to validate the forecast results by comparing historical electrical power variations against the predictions generated by the proposed model.

4.3.2. Calculation of the Avoidable Energy Bill

The avoidable energy bill was calculated using the following equation:

$${\mathrm{E}}_{\mathrm{b}\mathrm{i}\mathrm{l}\mathrm{l}}=\sum (T_{pk}·E_{pk})+\sum (T_{on}·E_{on})+\sum {(T}_{off}·E_{off})$$

(13)

where

• E_bill: the avoidable energy bill during a production cycle of two years by impeding the importation from the grid (USD);
• E_pk: avoidable energy during peak hours (MWh);
• E_on: avoidable energy during on-peak hours (MWh);
• E_off: avoidable energy during off-peak hours (MWh);
• T_pk: tariff of energy importation during peak hours (USD/MWh).
• T_on: tariff of energy importation during on-peak hours (USD/MWh);
• T_off: tariff of energy importation during off-peak hours (USD/MWh).

From a financial perspective, the energy imports totalling 7624.9 MWh, and recorded during the production cycle between 2024 and 2025, have caused costs of importations.

In other words,

Table 7. Bill of the imported energy from the grid according to tariff periods during 2024 and 2025.

Year	Imported Energy from the Grid (MWh)	Distribution of Imported Energy Across Tariff Periods (MWh)		Energy Tariff (USD/MWh)	Avoidable Energy Bill (USD)
2024	6138.8	Peak hours	1534.7	97.0	148,865.9	509,827.3
		On-peak hours	2578.3	85.0	219,155.1
		Off-peak hours	2025.8	70.0	141,806.2
2025	1486.1	Peak hours	371.53	97.0	36,037.9	123,420.6
		On-peak hours	624.16	85.0	53,053.7
		Off-peak hours	490.41	70.0	34,328.9

summarizes the avoidable energy bill through the use of the developed digital dashboard:

E_bill = 633,247.9 USD

4.3.3. Calculation of the Avoidable Carbon Emissions

Ultimately, the environmental impact is noticeable in the reduction in CO₂ emissions. In reality, the power plant under study is a cogeneration unit that emits no greenhouse gases. This means that avoiding energy imports also prevents emissions of this gas [35].

Knowing that this study was conducted in Morocco, the current carbon footprint in the country has seen a considerable decrease over recent years [36]. In 2025, for example, electricity production activity in Morocco generated approximately 607 g for each kWh produced from the entire energetic mix in the country, signifying specific carbon emissions of 0.607 tons/MWh.

Avoidable CO₂ emissions were calculated using the following equation:

$${\mathrm{T}}_{\mathrm{C}\mathrm{O}2}={\mathrm{E}}_{\mathrm{W}}·{\mathrm{S}}_{\mathrm{C}\mathrm{O}2}$$

(14)

where

• T_CO2: the avoidable CO₂ emissions during a production cycle of two years by impeding the importation from the grid (Ton);
• S_CO2: Current specific carbon emissions in Morocco (0.607 tons/MWh).

All calculations considered, the total carbon emissions for our case is about 4628 tons every two years, that could be avoided through the predictive decision-making solution that we have developed in the current paper.

5. Conclusions and Outlook

In the light of the above, this contribution has provided a tailor-made digital solution to help industrial leaders in improving the eco-efficiency of power stations based on the Rankine–Hirn cycle. The key to achieving this desired operational sustainability lies in increasing the responsiveness of decision-makers, by making quick and right decisions regarding the phenomenon of undesirable drops in energy production, thereby eliminating resource waste. The developed solution took the form of a sustainable digital dashboard, which was achieved using “Node-Red” and has included the three following interfaces:

The first interface was dedicated to the real-time display for tracking the behavior of various physical variables. This interface was equipped with an instantaneous notification system delivered at the following three levels: by local display of the potential deviation, through SMS, and by sending an email to the concerned parties.

The second interface provides information on the historical variation in all thirty-two physical parameters, allowing for the detection of those presenting incipient abnormal deviations and impacting the electric power production at the turbogenerator output.

The third interface features the results of a machine-learning-based predictive model, providing an anticipated insight into potential future variations in the electric power as the output variable of the model. Indeed, this model was developed using the XGBoost algorithm, whose learning results were satisfactory, with a determination score R² equal to 99.36% at the training stage and 99.11% at the test stage. Note that the choice of this algorithm was made following a comparative evaluation with the RFR and GRU algorithms.

Furthermore, by associating the developed dashboard with “Remote-Red” tool, we succeeded in creating a mobile version of the dashboard. This version aims to leverage the advantages provided by IoT devices, particularly real-time and mobility capabilities, which act as catalysts stimulating reactivity among decision-makers.

Later, the whole complementary digital solution was subjected to an experimental test using a simulation platform, which allowed for the prediction of a power drop 24 h in advance. This enabled the decision-maker to adjust the right physical variable causing the anomaly, thereby preventing the power drop phenomenon from occurring.

Moreover, through the analysis of the two-year historical production cycle, we were able to determine similar situations where production drops were avoidable. We found that the solution can help preserve up to 7.6 GWh of electricity every two years, which could be translated into a cost-saving exceeding 633,247.9 USD, and the avoidance of carbon emissions of 4628 tons of CO₂.

Furthermore, one of the foremost limitations of this study is the use of an inconsistent data history. Developing a predictive model based solely on the past two years of variations can lead to predictions with non-negligible margins of error in some cases. Consequently, the machine learning process will need to be repeated once a more substantial historical dataset is available.

As a perspective, the next step will be focused on the deployment of the solution in the real conditions of the process. Also, it is interesting to apply the same approach used in this work to other types of power plants, particularly those based on renewable energies such as thermo-solar power plants, where we find a significant potential of optimization in terms of energy efficiency and industrial decarbonization. The extent of this generalization is particularly significant for Concentrated Solar Power (CSP) facilities, which utilize the same Rankine–Hirn thermodynamic principles. In such contexts, the multi-interface dashboard could be seamlessly adapted to monitor solar-specific variables (e.g., DNI, molten salt temperatures) while maintaining the same predictive logic. Consequently, the proposed Decision Support System (DSS) serves not only as a solution for fossil-fuel plants but as a versatile tool for industrial decarbonization, capable of enhancing the operational reliability of any high-inertia thermal energy system.