Computer-implemented method for improving energy consumption of a cooling system

Abstract

The present invention relates to a computer-implemented method for improving energy consumption of a cooling system comprising one or more condensers and one or more compressors. The method comprises receiving previous sensor data associated with a cooling system, receiving a wet bulb temperature prediction, predicting a cooling system behaviour based on the previous sensor data and the wet bulb temperature prediction, using a trained machine learning model and adjusting one or more cooling system parameters based on the prediction. The invention also relates to a corresponding cooling system comprising means to carry out the method, to a method for training the machine learning model, to the trained machine learning model, and to a corresponding computer program.

Description

FIELD OF THE DISCLOSURE

[0001] The present disclosure relates to a system and method for improving the energy consumption of cooling systems comprising one or more condensers and compressors. Specifically, the present disclosure pertains to the use of machine learning models for predicting cooling system behavior and optimizing operational parameters based on sensor data and environmental conditions.BACKGROUND OF THE INVENTION

[0002] In the field of industrial and commercial cooling systems, efficient management of energy consumption remains a critical concern. Cooling systems, which typically comprise multiple compressors and condensers, are essential for maintaining desired temperature levels in various environments, such as data centers, manufacturing facilities, and large-scale refrigeration units. Known systems often rely on rule-based control strategies or static setpoints, which are adjusted based on limited sensor feedback or historical operational data. These conventional approaches may not fully account for dynamic environmental conditions, such as fluctuating ambient temperatures or varying humidity levels, which can significantly impact the performance and energy efficiency of cooling equipment. As a result, these systems may operate sub-optimally, leading to increased energy consumption, higher operational costs, and potential wear on critical components.

[0003] Despite advances in sensor technologies and the integration of basic automation in cooling systems, current methods for optimizing energy use are frequently limited by their inability to anticipate and adapt to changing external and internal conditions in real time. For example, many existing systems do not effectively utilize predictive information, such as weather forecasts or anticipated cooling loads, to proactively adjust operational parameters. Furthermore, the complexity of interactions among multiple system variables, such as compressor load, condenser fan speed, and refrigerant pressures, poses significant challenges for traditional control algorithms, which may not be capable of identifying the most energy-efficient operating points under varying circumstances. This often results in a reliance on conservative safety margins or fixed operational boundaries, further reducing the potential for energy savings.

[0004] It is therefore a technical problem underlying the present invention to provide improved methods and systems for optimizing the energy consumption of cooling systems that at least partially overcome the disadvantages of known systems.SUMMARY OF THE DISCLOSURE

[0005] It is an object of this invention to provide methods that overcome one or more of the disadvantages of known systems.

[0006] A first aspect of the invention provides a computer-implemented method for improving energy consumption of a cooling system comprising one or more condensers and one or more compressors, the method comprising: receiving previous sensor data associated with a cooling system; receiving a wet bulb temperature prediction; predicting a cooling system behavior based on the previous sensor data and the wet bulb temperature prediction, using a trained machine learning model; and adjusting one or more cooling system parameters based on the prediction.

[0007] The subject matter concerns a method executed by a computer for optimizing the energy consumption of a cooling system that includes at least one condenser and at least one compressor. The process begins with the acquisition of historical sensor data from the cooling system. This sensor data typically includes operational parameters such as pressures, temperatures, power consumption, and other relevant measurements collected from the system's components over a recent period, for example, the last two hours. Alongside this, a prediction of the wet bulb temperature is received. Wet bulb temperature is a meteorological parameter that reflects the lowest temperature air can reach by evaporative cooling and is highly relevant for cooling system efficiency, as it influences condenser performance and overall energy demand.

[0008] Using both the previous sensor data and the predicted wet bulb temperature, the method employs a trained machine learning model to forecast the future behavior of the cooling system. In this context, the machine learning model may function as a digital twin, capturing the real-world dynamics and interdependencies between system components such as compressors, condensers, and their control logic. The model is trained on historical plant data and is capable of simulating how the system will respond to different control setpoints and external conditions, rather than relying on idealized or theoretical assumptions.

[0009] Based on the prediction generated by the machine learning model, the method then adjusts one or more parameters of the cooling system. These parameters may include, for example, the discharge pressure setpoint of the compressors, which is a key lever for optimizing energy consumption. By dynamically selecting setpoints that minimize energy use while respecting operational constraints such as fixed suction pressure requirements, the method ensures that the system operates efficiently under varying real-world conditions.

[0010] The advantages of this approach are significant. By leveraging AI-driven modeling and real-time data, the method enables actionable and robust optimization of industrial cooling systems. It accounts for the actual operational behavior of the plant, adapts to changing environmental conditions, and continuously seeks to minimize energy consumption. This results in measurable reductions in power usage, improved cost efficiency, and enhanced sustainability, all while maintaining compliance with safety and performance constraints set by system operators or clients. The integration of predictive analytics and automated control adjustment represents a substantial improvement over static or rule-based control strategies, providing a competitive edge in the delivery of Cooling / Heating as a Service.

[0011] In a first implementation of the method according to the first aspect, the wet bulb temperature prediction is derived from the weather forecast.

[0012] This feature clarifies the source and temporal relevance of the wet bulb temperature data used in the method. The communication mechanism implied here may involve interfacing a cooling system's control logic, or the computer system implementing the method, with an external weather forecasting service or database. This may be achieved through network communication protocols such as API calls, data feeds, or other forms of electronic data exchange that enable the system to retrieve up-to-date weather forecast information. The new feature enhances the method by ensuring that the wet bulb temperature prediction is not merely a static or historical value, but is instead a dynamic, forward-looking parameter that reflects anticipated environmental conditions. By specifying that the prediction is derived from a weather forecast, the implementation ensures that the machine learning model's input is informed by real-time or near-future meteorological data, which can significantly improve the accuracy of the cooling system behavior prediction. The preference for using the forecast for the upcoming hour further refines the temporal granularity of the input, allowing the system to make more precise and timely adjustments to the cooling system parameters. This can lead to more responsive and efficient energy management, as the system can proactively adapt to imminent changes in ambient conditions rather than reacting to outdated or less relevant data.

[0013] In a further implementation of the method according to the first aspect, predicting the cooling system behavior comprises predicting a plurality of cooling system behaviors based on the previous sensor data and the wet bulb temperature prediction.

[0014] The implementation introduces a refinement to the prediction step by specifying that the prediction of cooling system behavior involves generating a plurality of predicted behaviors, each corresponding to a different potential adjustment of cooling system parameters. The mechanism of communication between the components in this context may involve the transmission of previous sensor data and wet bulb temperature prediction to the trained machine learning model, which then processes this input to output multiple predicted behaviors. Each predicted behavior may be associated with a specific parameter adjustment, which may imply that the model not only forecasts the system's response but also evaluates the impact of various possible adjustments. This enhances the system's ability to consider multiple operational scenarios in parallel, rather than limiting the prediction to a single outcome. By associating each predicted behavior with a corresponding adjustment, the method enables a more comprehensive analysis of potential control strategies, facilitating the selection of the most effective adjustment for optimizing energy consumption. This multi-scenario prediction capability provides a more robust decision-making framework, as it allows the system to weigh the potential outcomes of different parameter settings before implementing changes.

[0015] In a further implementation of the method according to the first aspect, further comprising: selecting the predicted cooling system behavior having a lowest predicted power consumption from the plurality of predicted cooling system behaviors; and adjusting the corresponding cooling system parameter accordingly.

[0016] The implementation introduces additional mechanisms and features that refine the operation of the method for improving energy consumption in a cooling system. Specifically, it describes a process where, from a plurality of predicted cooling system behaviors generated by the machine learning model, the method selects the behavior that is associated with the lowest predicted power consumption. This selection process implies that the system is capable of generating and evaluating multiple potential operational scenarios or configurations for the cooling system, each with its own predicted outcome regarding power usage. The mechanism of communication between components in this context may involve the transmission of predicted behaviors and their associated power consumption values from the machine learning model to a decision-making module or logic within the system. This module may be responsible for comparing the predicted power consumptions and identifying the optimal behavior, that is, the one with the lowest predicted energy use. Once this optimal behavior is selected, the system then may communicate the corresponding cooling system parameter adjustments to the control subsystem that manages the physical operation of the condensers and compressors. The above may be considered as optimization step, where the system does not merely predict a single behavior and adjust parameters accordingly, but rather evaluates multiple predicted behaviors and actively selects the one that minimizes energy consumption. This adds a layer of intelligence and efficiency to the method, ensuring that the adjustment of cooling system parameters is not just reactive to predictions, but is also optimized for energy savings. This optimization capability can lead to more significant reductions in energy usage and operational costs, as it allows the system to make informed decisions based on comparative analysis of predicted outcomes, rather than relying on a single predicted scenario. Thus, the implementation enhances the overall method by embedding a selection and optimization process that leverages the predictive capabilities of the machine learning model to achieve improved energy efficiency.

[0017] In a further implementation of the method according to the first aspect, further comprising determining a baseline power consumption of the cooling system.

[0018] The implementation introduces a mechanism for determining a baseline power consumption of the cooling system, which serves as a reference point for evaluating the effectiveness of predicted cooling system behaviors. The communication between components in this context may involve the transmission of sensor data from the cooling system to a computational module or server, where the baseline power consumption is calculated, such as through analysis of historical energy usage data under standard or typical operating conditions. The predicted cooling system behavior, generated by the trained machine learning model based on previous sensor data and wet bulb temperature prediction, may then be communicated to a comparison module or algorithm. This comparison module may receive both the predicted behavior and the established baseline power consumption, enabling a direct assessment of whether the predicted behavior represents an improvement or deviation from the norm.

[0019] In a further implementation of the method according to the first aspect, adjusting the one or more cooling system parameters further comprises: determining a boundary condition based on one or more of a predefined domain rule.

[0020] The implementation introduces additional mechanisms for how the adjustment of cooling system parameters is performed, specifically by incorporating the determination of a boundary condition. The communication between components in this context involves the system receiving or accessing information related to predefined domain rules, data-driven thresholds, and safety margins, which may be stored within the system or dynamically provided by external sources or subsystems. The trained machine learning model, after predicting the cooling system behavior, interacts with a decision-making module or logic that evaluates these boundary conditions. The determination of the boundary condition acts as an intermediary step, where the system processes input from domain knowledge, empirical data, or operational safety constraints to establish a permissible range or limit for parameter adjustment. This boundary condition is then communicated to the control or actuation module responsible for implementing the parameter changes in the cooling system. The new feature introduced by this implementation is the explicit integration of boundary conditions into the adjustment process, ensuring that any changes to the cooling system parameters are not only based on predictive analytics but are also constrained or guided by operational rules, learned thresholds, or safety considerations. This adds a layer of robustness and reliability to the system, as it prevents adjustments that could otherwise violate operational limits, compromise safety, or lead to suboptimal performance. By formalizing the use of boundary conditions derived from multiple sources, the implementation enhances the adaptability and safety of the control strategy, allowing the system to dynamically respond to both predicted behavior and real-world constraints. This feature distinguishes the method by embedding a structured decision framework that governs parameter changes, thereby improving the practical applicability and trustworthiness of the machine learning-driven control in real-world cooling systems.

[0021] In a further implementation of the method according to the first aspect, the predicted cooling system behavior includes a one or more of an expected fluid pressure, an expected condensing temperature, and / or an expected total power consumption value of the cooling system.

[0022] In a further implementation of the method according to the first aspect, the previous sensor data comprises one or more of a measured condensing temperature, a measured compressor driving power, a measured condenser fan speed, a saturated discharge temperature, a saturated suction temperature, an ambient air temperature, a condenser inlet temperature and / or a condenser outlet temperature.

[0023] In a further implementation of the method according to the first aspect, further comprising aggregating a plurality of previous sensor data points into one or more aggregated previous sensor data points, wherein the cooling system behavior is predicted based on the aggregated previous sensor data and the wet bulb temperature prediction, wherein preferably each type of previous sensor data is aggregated into a corresponding aggregated previous sensor data point. Alternatively, multiple types of sensor data are aggregated into a single aggregated previous sensor data point. For example, multiple pressures in the system and power consumptions may be averaged into a single aggregated previous sensor data point. In one example, multiple condenser discharge pressures and / or compressor cooling capacities may be aggregated into a single data point. Preferably, multiple condensers and / or compressors in individual stages are aggregated into a virtual condenser and / or compressor, wherein the sensor data of the respective condensers and / or compressors of the stage are aggregated into a single corresponding data point. This means that the virtual condenser and / or compressor is preferably resembled by a single corresponding data point. Preferably, the virtual condenser comprises at least 2, preferably at least 3, more preferably at least 4 and most preferably at least 5 condensers, and / or at most 20, preferably at most 15, more preferably at most 12 and most preferably at most 20 condensers. Preferably, the virtual compressor comprises at least 2, preferably at least 3, more preferably at least 4 and most preferably at least 5 compressors, and / or at most 20, preferably at most 15, more preferably at most 12 and most preferably at most 10 compressors.

[0024] The implementation introduces the feature of aggregating multiple previous sensor data points into one or more aggregated previous sensor data points prior to the prediction of cooling system behavior. The specific mechanism of communication between components in this context may involve the transfer of raw sensor data from the sensors associated with the cooling system to a data processing module or aggregation engine, which may be responsible for performing the aggregation operation. This aggregation engine may process the incoming stream or batch of sensor data, which may include various types such as temperature, pressure, flow rate, or other operational parameters, and combines them according to a defined aggregation function, such as averaging, summing, or extracting statistical features. The aggregated data points may then be communicated to the machine learning model, which receives both the aggregated sensor data and the wet bulb temperature prediction as inputs for the prediction of cooling system behavior. The new feature brought by this implementation is the transformation of raw, potentially high-frequency and high-volume sensor data into a more compact and information-rich representation before it is used for predictive modelling. This aggregation step can reduce noise, improve computational efficiency, and potentially enhance the robustness and accuracy of the machine learning model by focusing on salient trends or patterns rather than individual, possibly noisy, measurements.

[0025] In a further implementation of the method according to the first aspect, the previous sensor data includes sensor data obtained within the preceding 3h, preferably within the preceding 2.5h, more preferably within the preceding 2h.

[0026] The implementation introduces a temporal limitation on the previous sensor data utilized in the method, specifying that the sensor data must be obtained within a defined recent time window, such as within the preceding three hours, with preferences for even shorter intervals of two and a half hours or two hours. This feature refines the mechanism by which the system communicates and processes information between the sensors and the machine learning model. The sensors continuously or periodically collect operational data from the cooling system, and this data is transmitted to a processing unit or server where the machine learning model resides. By restricting the input data to only that which has been collected within a recent and specific timeframe, the system ensures that the predictions regarding cooling system behavior are based on the most current and relevant operational conditions. This temporal filtering can be implemented through a data management module or software logic that selects or queries only the sensor data entries that fall within the designated time window before providing them to the prediction module. The new feature thereby enhances the responsiveness and accuracy of the system by reducing the influence of outdated or less relevant data, which may not accurately reflect the present state or environmental conditions affecting the cooling system. As a result, the adjustments made to the cooling system parameters are more likely to be optimal for the current or imminent conditions, leading to improved energy consumption outcomes. This temporal specificity also facilitates more dynamic adaptation to rapidly changing environmental factors, such as fluctuations in wet bulb temperature, by ensuring that the predictive model is always operating on the freshest available data.

[0027] In a further implementation of the method according to the first aspect, intervals of at least 30 minutes, preferably at least 45 minutes and most preferably at least 60 minutes, and / or at most 120 minutes, preferably at most 90 minutes and most preferably at most 60 minutes are aggregated into a single data point corresponding to the respective type of sensor, preferably by averaging the corresponding sensor data over the duration of interval.

[0028] The implementation introduces a specific data aggregation mechanism that operates over defined time intervals, which are set to be at least 30 minutes and at most 120 minutes, with preferred and most preferred ranges further narrowing these bounds. This aggregation may be performed by averaging the sensor data collected over the specified interval, resulting in a single representative data point for each sensor type. The mechanism of communication between the system components may involve the transmission of raw sensor data from the cooling system sensors to a data processing module, which may then perform the aggregation operation. This aggregated data point may subsequently be communicated to the machine learning model or prediction engine, which utilizes it in conjunction with other inputs, such as wet bulb temperature predictions, to forecast cooling system behavior. By averaging over longer intervals, the system can achieve a more stable and representative input for predictive modeling, which may enhance the reliability and accuracy of subsequent predictions and adjustments to cooling system parameters. Furthermore, the defined interval ranges allow for flexibility in system configuration, enabling adaptation to different operational requirements or environmental conditions. This feature also facilitates more efficient data storage and transmission, as fewer data points need to be managed and communicated between system components, thereby optimizing computational resources and bandwidth.

[0029] In a further implementation of the method according to the first aspect, the cooling system parameter includes one or more of a compressor power supply, a set condensing temperature, a condenser fan speed, an expansion valve setting, a compressor driving power, a suction temperature, a discharge pressure, a discharge temperature or combinations thereof.

[0030] The implementation introduces a refinement to the method by specifying that the cooling system parameter, which is subject to adjustment based on the machine learning prediction, encompasses a defined set of operational variables. The specific mechanisms of communication between components implied by this implementation may involve the transmission of control signals or commands from a central processing unit or controller, which executes the machine learning-based prediction, to the respective actuators or control modules associated with each of the enumerated parameters. For example, the controller may send a signal to adjust the compressor power supply, modulate the set condensing temperature, or vary the speed of the condenser fan, each of which would require an established communication protocol, such as analog or digital control lines, fieldbus communication, or wireless signals, between the controller and the relevant hardware component.

[0031] A second aspect of the invention provides a cooling system comprising one or more condensers; one or more compressors; and means for carrying the computer-implemented method according to the preceding aspect.

[0032] The subject matter described is a cooling system that integrates both traditional hardware components and advanced digital control. The system includes one or more condensers, which are devices responsible for condensing refrigerant vapor into liquid, thereby releasing heat from the system. It also comprises one or more compressors, which are mechanical devices that increase the pressure of the refrigerant, enabling the refrigeration cycle to function efficiently by moving the refrigerant through the system and maintaining the necessary pressure differentials.

[0033] In addition to these physical components, the system incorporates means for carrying out a computer-implemented method as previously defined. This refers to the integration of an AI-driven optimization process, specifically leveraging a digital twin model, such as a Vector Autoregressive (VAR) model. This digital twin may be trained using historical plant data and external variables like wet bulb temperature, allowing it to accurately simulate real-world system dynamics rather than relying on idealized or theoretical models. The AI model may forecast future system behavior, including pressures, capacities, and energy consumption, for various control setpoints, particularly focusing on optimizing the discharge pressure of the compressors.

[0034] The primary advantage of this configuration lies in its ability to dynamically adjust operational parameters in real time to minimize overall energy consumption, while still adhering to operational constraints such as fixed suction pressure requirements. By continuously simulating and selecting the optimal setpoints based on recent operational data and weather forecasts, the system achieves significant energy savings and operational efficiency. The integration of the AI-based digital twin ensures that optimization decisions are robust, data-driven, and tailored to actual plant conditions, resulting in measurable reductions in power consumption compared to traditional baseline operations. This approach enables a seamless bridge between domain expertise and advanced AI, enhancing the overall performance and sustainability of industrial cooling systems.

[0035] A third aspect of the invention provides a computer-implemented method for training a machine learning model configured to predict a cooling system behavior of a cooling system comprising one or more condensers and one or more compressors, the method comprising obtaining training data from one or more cooling systems comprising a plurality of parameter sets comprising one or more of: a discharge temperature from the compressor, a discharge pressure from the compressor, a suction pressure from the compressor, a suction temperature from the compressor, a power consumption of the compressor, a power consumption of the condenser, a fan speed of the condenser, a fan power consumption of the condenser, a temperature at the condenser input, a temperature at the condenser output, an expansion valve position, a wet bulb temperature, an ambient air temperature, a cooling load demand, a time of day, seasonal pattern; training the machine learning model based on the training data to generate a trained machine learning model.

[0036] The subject matter concerns a method executed by a computer for training a machine learning model that is designed to predict how a cooling system will behave. This cooling system includes at least one condenser and at least one compressor, which are standard components in industrial refrigeration or air conditioning systems. The method involves collecting training data from one or more such cooling systems. The collected data consists of multiple sets of parameters, where each set may include values such as the temperature and pressure at both the discharge and suction sides of the compressor, the power consumption of both the compressor and the condenser, the speed and power consumption of the condenser fan, temperatures at the condenser's input and output, the position of the expansion valve, the wet bulb temperature (which reflects both temperature and humidity of the air), the ambient air temperature, the current cooling load demand, the time of day, and any seasonal patterns.

[0037] The method then uses this training data to train the machine learning model, resulting in a trained model that can predict the behavior of the cooling system under various conditions. In this context, "training data" refers to historical or real-time operational data gathered from actual cooling systems, capturing how different parameters interact and affect system performance. The "machine learning model" is an algorithmic construct, such as a vector autoregressive model, that learns from this data to recognize patterns and relationships among the parameters.

[0038] By using a wide range of operational parameters, the method ensures that the model captures the complex interdependencies between system components, environmental conditions, and control settings. This approach moves beyond theoretical or idealized models by reflecting real-world system dynamics, including the effects of weather, time, and varying load demands. The advantage of this subject matter is that it enables more accurate and actionable predictions of cooling system behavior, which can be used to optimize control strategies, such as dynamically adjusting discharge pressure to minimize energy consumption while maintaining required performance constraints. This leads to measurable efficiency gains, reduced operational costs, and improved integration of domain knowledge with artificial intelligence in industrial cooling applications.

[0039] A fourth aspect of the invention provides a trained machine learning model configured to predict a cooling system behavior of a cooling system comprising one or more condensers and one or more compressors, wherein the trained machine learning model is trained according to the preceding aspect.

[0040] The subject matter described is a trained machine learning model specifically configured to predict the behavior of a cooling system that includes one or more condensers and one or more compressors. The model is not generic but is trained according to a particular methodology that has been previously defined, which involves a multi-step process leveraging historical plant data and external inputs, such as wet bulb temperature, to reflect real-world operational dynamics rather than theoretical or idealized system behavior.

[0041] A machine learning model in this context refers to a computational algorithm that has been exposed to large amounts of historical operational data from actual cooling systems. Through this training, the model learns the complex interdependencies and dynamic responses of the system components, such as how changes in discharge pressure, suction pressure, and ambient conditions affect overall energy consumption and system performance. The model configuration is tailored to the architecture of industrial cooling systems, which typically feature multiple condensers and compressors arranged in high and low temperature circuits.

[0042] The prediction of cooling system behavior encompasses forecasting key operational parameters, including pressures, capacities, and energy consumption, under various control setpoints and anticipated external conditions. The model is trained using a method that first derives valid ranges for control variables based on domain-specific rules and safety margins, then simulates system responses for each candidate setpoint using a digital twin approach and finally selects the setpoint that minimizes energy use over a defined period, such as 2 hours, while respecting operational constraints set by clients.

[0043] The advantage of this approach lies in its ability to provide actionable forecasts and optimization recommendations that are grounded in actual plant behavior, rather than relying on static models or theoretical assumptions. By continuously retraining and updating the model with new operational data, the system adapts to evolving plant conditions and maintains high accuracy in its predictions. This enables dynamic adjustment of control setpoints, such as discharge pressure, to achieve significant reductions in energy consumption without compromising system safety or client-defined constraints. The model also facilitates fair baseline comparisons by predicting expected power consumption based on historical data, allowing for clear measurement of efficiency gains resulting from optimization.

[0044] In a first implementation of the trained machine learning model of the preceding claim according to the fourth aspect, the machine learning model is one or more of a Long Short-Term Memory, LSTM, Network, a Recurrent Neural Network, RNN, a Vector Error Correction Model, VECM, a Support Vector Regression, SVR, model, a Physics-Informed Neural Network, PINN, a Deep Q-Networks, DQN, or a Vector Auto Regression, VAR, model.

[0045] The implementation introduces a specific set of machine learning model architectures that may be employed to predict the cooling system behavior, thereby refining the scope of the aspect by defining the nature of the trained machine learning model. The mechanisms of communication between components in this context pertain to the internal structure and operational dynamics of the specified models. For instance, a Long Short-Term Memory (LSTM) network and a Recurrent Neural Network (RNN) both utilize feedback connections, allowing information to persist and flow across time steps, which is particularly advantageous for capturing temporal dependencies in sequential data, such as the time-series data generated by cooling system sensors. A Vector Error Correction Model (VECM) facilitates communication between multiple time series by modeling both their short-term dynamics and long-term equilibrium relationships, enabling the model to account for interdependencies among variables such as compressor pressure and condenser temperature. The Support Vector Regression (SVR) model operates by mapping input features into a high-dimensional space and finding the optimal hyperplane for regression, where communication between components is realized through kernel functions that relate input data points to the support vectors. Physics-Informed Neural Networks (PINNs) integrate domain-specific physical laws directly into the learning process, ensuring that the model's predictions are consistent with known physical relationships, thereby establishing a form of communication between empirical data and theoretical models. Deep Q-Networks (DQNs) employ a reinforcement learning framework where the model communicates with its environment by receiving state information, selecting actions, and updating its policy based on observed rewards, which is suitable for optimizing control strategies in dynamic cooling systems. Vector Auto Regression (VAR) models capture linear interdependencies among multiple time series, allowing for simultaneous modeling of several system variables and their mutual influences.

[0046] A fifth aspect of the invention provides a computer program comprising instructions which, when executed by a processor, cause the processor to carry out the computer-implemented method according to the first aspects or any one of its implementations. Preferably, the fifth aspect provides a non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method of the first aspect and / or corresponding implementations.

[0047] The subject matter described involves a computer program that, when executed by a processor, implements a method for optimizing the operation of industrial cooling or heating systems using artificial intelligence. The computer program contains instructions that enable the processor to execute a series of steps, including the derivation of valid control variable boundaries, simulation of system behavior using a digital twin model based on a vector autoregressive approach, and selection of optimal setpoints for system parameters such as discharge pressure. Preferably, the instructions are stored on a non-transitory computer-readable medium that, when executed by one or more processors, cause the one or more processors to perform the steps. The program utilizes recent historical plant data and predicted external conditions, such as wet bulb temperature, to forecast system performance under various scenarios. By evaluating these scenarios, the program identifies the setpoint configuration that minimizes energy consumption over a defined period, typically 2 hours, while ensuring that operational constraints, such as fixed suction pressure limits, are maintained. The program also facilitates the calculation of baseline energy consumption for comparison, using either a model-based prediction or a similarity search in historical data, with normalization of feature ranges to ensure accurate comparisons. Integration with plant control systems allows the program to automatically update operational setpoints in real time, leveraging ongoing data collection and periodic retraining of the underlying AI model to maintain accuracy and efficiency. The advantages of this approach include significant reductions in energy usage, improved alignment of system operation with real-world conditions, and the ability to continuously adapt to changing plant dynamics and environmental factors, thereby enhancing the overall efficiency and reliability of cooling or heating as a service offerings.BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1:is a schematic block diagram of an embodiment of a cooling system; Figure 2:is a schematic block diagram of an embodiment of a two-stage cooling system; Figure 3:is a schematic flowchart illustrating an embodiment of a method for improving energy consumption of a cooling system; Figure 4:is a graph showing system behavior based on adjusted system parameters according to an embodiment. DETAILED DESCRIPTION

[0049] Figure 1 shows a schematic representation of a cooling system 100 according to an embodiment. The cooling system 100 comprising a condenser 110, a compressor 120, a cooler 130, and an expansion valve 140. Preferably the condenser 110, the compressor 120, the cooler 130, and the expansion valve 140 are interconnected in a closed-loop circuit. The system is configured for the circulation and thermodynamic manipulation of a working fluid, typically a refrigerant, to achieve heat transfer and cooling.

[0050] The condenser 110 is positioned downstream of the compressor 120 and upstream of the expansion valve 140. In operation, the condenser 110 receives high-pressure, high-temperature vaporized refrigerant discharged from the compressor 120. Preferably, the condenser 110 is configured to provide heat rejection from the refrigerant to the ambient environment, preferably via forced or natural convection, resulting in condensation of the vapor into a high-pressure liquid. Preferably, the condenser 110 is controlled by adjusting one or more system parameters associated with the condenser. Exemplary system parameters associated with the condenser include condensing temperature, condenser inlet temperature, condenser outlet temperature, condenser fan speed, and condenser power provision. Based on the condensing temperature, the efficiency of heat rejection and the overall system coefficient of performance COP may be determined. Wherein the COP cooling = T l / (T h - T l ), where T l = cooler (low-side) temperature in Kelvin T h = condenser (high-side) temperature in Kelvin. Preferably the condenser 110 is in contact with ambient air or an ambient liquid.

[0051] Since the condenser is configured to perform heat exchange between the cooling system and ambient environment, the cooling system may operate under the influence of various environmental parameters which may affect the efficiency and capacity of the cooling system. The environmental parameter may be one or more of ambient air temperature, wet bulb temperature, time of day, and seasonal pattern. Preferably, the environmental parameter is the wet bulb temperature. The condenser system parameter may be adjusted based on one or more environmental parameters.

[0052] The compressor 120 is located downstream of cooler 130 and upstream of the condenser 110. The compressor 120 is configured to intake low-pressure vaporized refrigerant from cooler 130 and compresses it to a higher pressure and temperature, enabling subsequent heat rejection in condenser 110. The compressor 120 may be driven by an electric motor. Preferably, the compressor 120 operation is controlled by adjusting one or more system parameters associated with the compressor. Exemplary system parameters associated with the compressor 120 include a compressor driving power, a discharge pressure, a discharge temperature, a suction pressure, a suction temperature, and a compressor power supply. The discharge pressure and temperature may be indicative of the thermodynamic state of the refrigerant post-compression, while suction pressure and temperature may reflect the conditions at the compressor inlet.

[0053] The cooler 130 is situated downstream of expansion valve 140 and upstream of the compressor 120. The cooler 130 receives low-pressure liquid refrigerant from the expansion valve 140 and may facilitate heat absorption from the target environment or process, resulting in vaporization of the refrigerant. The cooler 130 may be implemented as an evaporator, where heat exchange occurs between the refrigerant and the medium to be cooled. System parameters relevant to the cooler 130 include a cooling load demand, an evaporator inlet temperature, an evaporator outlet temperature, and a fluid flow rate. The cooling load demand may quantify the thermal energy to be removed, while inlet and outlet temperatures may provide insight into the effectiveness of heat absorption.

[0054] The expansion valve 140 is arranged between the condenser 110 and the cooler 130. The expansion valve 140 may be configured to regulate the flow and pressure drop of the refrigerant entering cooler 130, enabling controlled vaporization and heat absorption. The expansion valve 140 may be implemented as a thermostatic expansion valve, electronic expansion valve, or other suitable metering device. Parameters associated with expansion valve 140 may include an expansion valve position, a flow rate, and a pressure differential across the valve. The expansion valve position may be modulated in response to the system demand and operating conditions to optimize refrigerant flow and maintain desired superheat levels.

[0055] The closed-loop configuration depicted in Figure 1 enables continuous circulation of the refrigerant through the four principal components, with each component performing a distinct thermodynamic function.

[0056] While Fig. 1 describes the cooling system as a single stage cooling system, multiple cooling stages may also be provided. For example, the cooling system described with respect to Fig. 1 may be extended to a multi-stage cooling system, such as a two-stage, three-stage, four-stage or five-stage cooling system.

[0057] For example, a two-stage multi cooling system may comprise of two of the circuits shown Fig. 1, wherein the cooler associated with the high stage 300 refrigeration circuit may be replaced with a heat exchanger 350 and the condenser associated with the low stage 200 refrigeration circuit may be replaced with a heat exchanger 250. In this case, instead of rejecting heat to ambient, the low stage heat exchanger rejects heat into the high stage 300 refrigeration circuit, by means of the heat exchanger 350 of the high stage 300 heat exchanger. This way, the low stage 200 refrigeration circuit can be cooled to significantly lower temperatures, while low temperatures can be maintained more stable. In particular, since the low stage 200 refrigeration circuit is not couples to ambient. Preferably, the low stage 200 refrigeration circuit provides a temperature as low as -20°C to - 40°C. Preferably, the high stage 300 refrigeration circuit provides a temperature

[0058] By further adding additional stages with corresponding heat exchangers, the above described two-stage cooling system can be enhanced to a three-stage cooling system, a four-stage cooling system etc.

[0059] Figure 2 illustrates an embodiment of a two-stage multi-stage cooling system. Any aspects described with respect to Fig. 1 may apply to the embodiment of Fig. 2. The two-stage cooling system of Fig. 2 comprises a high-stage 300 and a low-stage 200 refrigeration circuits. Each of the high-stage 300 and low stage 200 refrigeration circuits are organized into individual loops, each incorporating one or more compressors, heat exchangers, and flow control valves.

[0060] The lower section of the figure shows the low-stage 200 refrigeration circuit. This circuit includes a cooler 230, which is configured to receive and cool the working fluid / system. Downstream of the cooler 230, the fluid flow is directed toward one or more low-stage compressors 220. In the example shown in Fig. 2, three low-stage compressors 220 LS Compressor 1, LS Compressor 2, and LS Compressor 3 are provided. However, the invention is not limited to this specific number. Preferably, the one or more low-stage compressors 220 are arranged in parallel, allowing for staged or load-sharing operation. The compressed fluid from the one or more low-stage compressors 220 is then routed to one or more low-stage heat exchangers 250. One or more low-stage expansion valves 240 may be positioned upstream of the one more low-stage heat exchangers 250. The one or more low-stage expansion valves 240 may be configured to control the distribution and direction of the fluid in the low stage refrigeration circuit.

[0061] The one or more heat exchangers 250 serve as interface between the low-stage 200 and high-stage 300 refrigerator circuits, facilitating thermal exchange and further conditioning of the refrigerant.

[0062] The (heat) output from the one or more heat exchangers 250 is then transferred to one or more high stage heat exchangers 350. The one or more heat exchangers 350 are configured to receive heat from the low stage 200 and transfer the heat to the high stage 300.

[0063] Preferably, the one or more heat exchangers 250 and the one or more heat exchangers 350 are in fluid connection with a first refrigerant liquid provided in the low-stage cooling circuit. Preferably, the one or more low-stage heat exchangers 250 are not in fluid connection with a second refrigerant liquid provided in the high-stage cooling circuit 300. Preferably, the first and / or second refrigerant liquids are one or more of a natural refrigerant, such as R-717, R-744, R-600a or R-290, hydrofluorocarbon, HFC, such as R-134a, R-410A or R-32, a hydrofluorolefin, HFO, such as R-1234yf, R-1234ze. Preferably, the first and / or second refrigerant liquid is a natural refrigerant, such as R-717, R-744, R-600a or R-290.

[0064] From the one or more high-stage heat exchangers 350, the heat is directed to one or more high-stage compressors 320, by means of the second refrigerant liquid. The second refrigerant liquid is provided only in the high (or second) stage of the cooling system. In the example shown in Fig. 2, the high-stage 300 cooling circuit comprises two high-stage compressors 320, labelled HS Compressor 1 and HS Compressor 2. HS Compressor 1 and HS Compressor 2 are preferably also arranged in parallel. These high-stage compressors 320 may further elevate the pressure and temperature of the second refrigerant liquid.

[0065] Following compression in the one or more high-stage compressors 320, the fluid is conveyed to one or more high-stage condensers 310. The one or more high-stage condensers 310 are preferably configured in parallel, enabling efficient heat rejection from the high-pressure refrigerant to the ambient environment. In a three-stage (or higher stage) refrigerant circuit, the one or more high-stage condensers 310 are configured to reject the heat from the high-stage 300 cooling circuit to a higher-stage cooling circuit (not shown).

[0066] The condensed second refrigerant liquid is then routed through one or more high-stage expansion valves 340, which manage the return flow of the second refrigerant liquid toward the one or more heat exchangers 350 in the high-stage 300 cooling circuit.

[0067] The system architecture as depicted in Figure 2 demonstrates a cascaded refrigeration arrangement, where the low-stage and high-stage circuits are thermally coupled via heat exchangers 250, 350. The inclusion of multiple compressors in both stages allows for operational flexibility, redundancy, and load balancing. The use of flow control valves 240, 340 ensures precise management of refrigerant flow between the respective circuit components, optimizing system performance and energy efficiency.

[0068] The interconnection of the cooler 230, low-stage compressors 220, heat exchangers 250, high-stage compressors 320, condensers 310, and flow control valves 240, 340 establishes a closed-loop system capable of delivering staged cooling with high reliability and adaptability to varying load conditions. The parallel configuration of compressors and condensers further enhances the system's ability to modulate capacity and maintain stable operation across a range of thermal demands.

[0069] Figure 3 illustrates a flowchart representing a computer-implemented method for optimizing energy consumption in a cooling system according to an embodiment.

[0070] In a first step S410, previous sensor data associated with a cooling system is obtained. Preferably, the cooling system is any of the above-described cooling systems. Thus, any aspects described above with respect to the cooling systems may apply to the method described herein.

[0071] The obtained previous sensor data may comprise one or more of a measured condensing temperature, a measured compressor driving power, a measured condenser fan speed, a saturated discharge temperature, a saturated suction temperature, an ambient air temperature, a condenser inlet temperature and / or a condenser outlet temperature. The data is typically collected from sensors positioned at strategic locations within the cooling system, such as at the compressor, condenser, expansion valve, and ambient environment.

[0072] Preferably, the previous sensor data is collected for a certain period of time. For example, the previous sensor data may be collected over the past 4 hours, the past three hours, the past two hours, the past hour or over the past 30 minutes. Preferably, the previous sensor data is collected over the past two hours. Preferably, each sensor reading within the certain period of time is associated with one or more data points.

[0073] Preferably, a plurality of previous sensor data points are aggregated into one or more aggregated previous sensor data points. Preferably each type of previous sensor data is aggregated into a corresponding aggregated previous sensor data point. Preferably, intervals of at least 30 minutes, preferably at least 45 minutes and most preferably at least 60 minutes, and / or at most 120 minutes, preferably at most 90 minutes and most preferably at most 60 minutes are aggregated into a single data point corresponding the respective type of sensor. Preferably, the data points are aggregated by forming the average of the respective values over the interval. For example, in the example of a two hour aggregation of condenser discharge pressure, each hour of condenser discharge pressure measurement may be averaged into a single aggregated discharge pressure data point. Thereby, two aggregated condenser discharge pressure data points are obtained.

[0074] The collecting and / or averaging of individual data points may be performed before the previous sensor data is obtained in step S410, or after the previous sensor data is obtained in step S410. If the collecting and / or averaging of the individual data points is performed after the previous sensor data is obtained in step S410, then the individual data points are obtained in step S410. Otherwise, only the aggregated (such as two aggregated) data points are obtained in step S410. Preferably, two aggregated data points corresponding to the past two hours of cooling system operation are obtained in step S410.

[0075] In step S420, a wet bulb temperature prediction is obtained. The wet bulb temperature may be obtained by simulating a wet bulb prediction based on historic wet bulb temperature sensor measurements. Preferably, the wet bulb temperature prediction is derived from the weather forecast, preferably for the upcoming hour.

[0076] In step S430, a cooling system behavior is predicted based on the previous sensor data and the wet bulb temperature prediction. In this step, the previous sensor data and the wet bulb temperature prediction are processed using a trained machine learning model. The prediction is generated by correlating the historical sensor data and environmental conditions to anticipated system responses under similar circumstances. Preferably, the cooling system behavior is predicted for the upcoming hour, preferably using the predicted wet bulb temperature for the upcoming hour.

[0077] Preferably, the predicted cooling system behavior includes one or more of an expected fluid pressure, an expected condensing temperature, and / or an expected total power consumption value of the cooling system. Preferably, a plurality of cooling system behaviors are predicted based on the previous sensor data and the wet bulb temperature prediction, wherein each predicted cooling system behavior of the plurality of predicted cooling system behaviors is associated with a corresponding cooling system parameter adjustment. For example, a cooling system adjustment may be a change of the condenser discharge pressure. In this case, each predicted cooling system behavior of the plurality of predicted cooling system behaviors may be associated with a respective condenser discharge pressure change. For example, if 10 condenser discharge pressure changes are used as inputs, 10 corresponding predicted cooling system behaviors may be output from the trained machine learning model.

[0078] Based on the prediction, one or more cooling system parameters are adjusted. In the above example, based on the predicted cooling system behaviors, the corresponding discharge pressure may be changed accordingly.

[0079] Preferably, adjustment of the one or more cooling system parameters is based on selecting the predicted cooling system behavior having a lowest predicted power consumption from the plurality of predicted cooling system behaviors and adjusting the corresponding cooling system parameter accordingly. Preferably, each of the predicted power consumption of the cooling system includes the total power system of the entire cooling system. For example, the predicted power consumption of the cooling system may include the predicted power consumption of the condenser, of the compressor and of the extension valve. Preferably, the trained machine learning model is configured to predict each of the respective power consumptions based on (only) the obtained sensor data, such as on the obtained aggregated previous sensor data. In a preferred example, the trained machine learning model is able to predict the cooling system power consumption based on the expected wet bulb temperature and two aggregated previous sensor data points associated with the condenser discharge pressure.

[0080] The above may be performed for a single cooling system parameter type (such as only the condenser discharge pressure), or multiple cooling system parameter types (such as also the condenser power provision).

[0081] Preferably, selecting the predicted cooling system behavior includes comparing the selected predicted cooling system behavior with a baseline power consumption of the system. The baseline power consumption may indicate the predicted cooling system behavior without changes to one or more cooling system parameters. The predicted cooling system behavior that has a lower power consumption and has the most difference in power consumption when compared to the baseline power consumption, may be selected as the predicted cooling system behavior having a lowest predicted power consumption. Preferably, the baseline power consumption is determined using the trained AI model.

[0082] For example, during training of the AI model, the model may also learn typical cooling system of the past. The trained model may then be used to predict all cooling system parameters including suction pressure and consumed power, based on the previous sensor data and the expected wet bulb temperature. The sum of the consumed power of all components of the cooling system may then the calculated as the baseline power consumption against which the cooling system parameter adjustments are compared against. This means, during inference, the trained AI model determines the baseline power consumption for the cooling system based on only the previous sensor data and the expected wet bulb temperature and compares the determined baseline power consumption to the predicted system behavior including the corresponding amended system parameters. Based on this comparison, the lowest predicted power consumption may be derived.

[0083] Additionally, or alternatively, a historic baseline may be determined. For example, the predicted wet bulb temperature and a predicted compressor capacity are used to query the N most similar datapoints in the historic data. The value of N of the N most similar datapoints in the historic data may depend on the amount of available historic data. For example, if the historic data spans a total duration of only a few months, N may be in the range of ~ 1-10. If the historic data spans a total duration of 1 to 2 years, N may be in the range of ~ 50-100. The average total power consumption of the N data points is used as historic baseline. In addition to the compressor capacity other features may also be used to determine the most similar datapoint, such as the suction pressure on the different stages, etc. Preferably, a consequent normalization of the data is maintained, where the data range of each feature is mapped on a normalized interval, e.g. [0,1], before distances in the n-dimensional space are calculated.

[0084] After the cooling system behavior is predicted, in step S440, one or more cooling system parameters are adjusted based on the prediction. The adjustments may include altering one or more of a compressor power supply, a set condensing temperature, a condenser fan speed, an expansion valve setting, a compressor driving power, a suction temperature, a discharge pressure, a discharge temperature or combinations thereof. Preferably, the adjustment includes adjusting the condenser discharge pressure. The adjustments may be implemented in real time or near-real time.

[0085] Preferably, step S440 further comprises determining a boundary condition based on one or more of a predefined domain rule, a data-driven threshold and / or a safety margin, wherein the one or more cooling system parameters are adjusted based on the boundary condition. This may mean that the one or more cooling system parameters are only adjusted within the margins of the predefined domain rule.

[0086] While the above is described with respect to a single (single stage or multi-stage) cooling system, the above may also be implemented for a plurality of cooling systems. For example, the obtained previous sensor data may be a plurality of previous sensor data associated with a plurality of cooling systems. Likewise, the wet bulb temperature prediction may be a plurality of wet bulb temperature predictions, each wet bulb temperature prediction of the plurality of wet bulb temperature predictions may be associated with a corresponding cooling system of the plurality of cooling systems. The data obtained from each of the cooling systems of the plurality of cooling systems may be averaged into a single data point. Based on the averaged (aggregated) data points of the plurality of cooling systems, the above-described steps S430 and S440 may be performed. For example, the trained machine learning model may determine the predicted cooling system behavior and adjust a cooling system parameter based on the averaged data points of the plurality of cooling systems. In such a case, in each of the cooling systems of the plurality of cooling systems the same cooling system parameter may be changed. The plurality of cooling systems may also be referred to as a virtual cooling system.

[0087] Figure 4 shows two exemplary implementation results [a] and [b] obtained by applying the above described method to an exemplary cooling system. On the y-axis, the power consumption (absorbed power) in kW is shown. The x-axis corresponds to the time. The circular points / dots correspond to the system power consumption, based on the above referenced methods. The square-shaped points / dots correspond to the baseline system power consumption. The points / dots indicate the power consumption per hour.

[0088] Fig. 4 [a] shows a total timespan of 48 hours. The power consumption is shown from 0 kW to above 200 kW. Each column in the x-direction corresponds to a timespan of 6h. As can be seen from example [a] in Fig. 4, the power consumption based on adjusted system parameters was always below the power consumption of the baseline, i.e., of the cooling system without adjusted parameters according to the above methods.

[0089] Example [b] shows similar results over a total timespan of 10 days. Each column corresponds to timespan of 24h. While there are a few exceptions, almost all power consumption values were lower for the adjusted system parameters. In fact, with the above-described method, the total power consumption of the cooling system was reduced by about 10% (4.108kW of a total of 47.621 kW).

[0090] An exemplary training pipeline of a core machine learning model for cooling system optimization may comprise among others a step of data collection, a step of data preparation, a step of model training / model calibration and a step of model running. In addition, there may be a model building step, a model export step and a step of presenting prediction results on a customer interface.

[0091] The data collection step and data preparation step may comprise collecting the operational parameters over a period of time to obtain samples and, in case of supervised learning, labeling the collected sample (i.e., cooling system operating optimally vs. cooling system requiring efficiency improvement). Preferably, the period of time to obtain samples is at least a month, preferably a year. Preferably, the training data is provided in 3h buckets, comprising 2h of previous sensor data and corresponding 1h of corresponding predicted wet bulb temperature. An example would be the measurement of condenser discharge pressure and / or compressor power consumption during operational cycles, as discussed in more detail in Fig. 1 and Fig. 2. Additional operational parameters such as discharge temperature from the compressor, discharge pressure from the compressor, suction pressure from the compressor, suction temperature from the compressor, compressor cooling capacity, power consumption of the compressor, power consumption of the condenser, fan speed of the condenser, fan power consumption of the condenser, temperature at the condenser input, temperature at the condenser output, an expansion valve position, a wet bulb temperature, an ambient air temperature, a cooling load demand, a time of day and or seasonal pattern, which may be obtained using sensor measurements may also serve as input parameters to the machine learning model. Additional sensor readings may improve robustness of the model; however, it may be advantageous to limit the input parameter to input parameters that are available by default (i.e., no retrofitting of the industrial cooling system). Another option may be to use simulated data as input data to the machine learning model. Such simulated data may be generated by using a test cooling system. Moreover, a plurality of samples may be collected from a plurality of types of cooling systems which has the advantage of improving generalizability and accuracy of the prediction. In addition, the plurality of samples may be generated over a longer period of time from at least one industrial cooling system in an operational environment to represent the natural degradation of system efficiency. Another part of the data preparation step may comprise pre-filtering of the input data using visualization techniques such as heatmaps. Moreover, a pre-screening of the data may be performed using statistical values (i.e., average) of the input data and comprises grouping and further examining samples that qualify as outliers. Moreover, a trained professional may analyze the operational parameters for anomalies. For instance, the professional might check pressure fluctuations, temperature variations and power consumption peaks for condenser discharge pressure and compressor power consumption. The professional might also perform crosschecking the time stamp of the operational data profile with auxiliary sensor values such as ambient temperature and cooling load to accurately label data.

[0092] Feature extraction and data splitting may precede the training step. Feature extraction may comprise the identification and selection of relevant features which may be based on statistical values such as minimum, maximum, count and mean or specific characteristics that are defined by trained professionals. It is to note that feature extraction can also be performed using a dedicated machine learning model such as a convolutional neural network. Data splitting, which refers to dividing the data for training, testing and validation, may also be applied. During the training step two types of problems can be solved, a primal problem which aims at minimizing the optimization error (e.g., using a loss function for energy efficiency) or a dual problem which also aims at minimizing the optimization error but does so while maximizing an objective function using the Lagrange multipliers. Finally, an efficiency threshold is set that determines which samples are considered optimal and which samples are considered to be suboptimal. The threshold may be based on the samples distance from the decision boundary. Finally, the performance of the model may be evaluated using the validation data set and metrics such as precision, recall, F1-score, or area under the ROC curve (AUC). Moreover, the model's generalization capabilities and potential issues such as overfitting or underfitting may be identified.

[0093] The retraining step may allow the trained machine learning model to be scaled and applied to other cooling applications (e.g., industrial refrigeration, HVAC systems, process cooling) and / or cooling system types (e.g., varying compressor capacity, different refrigerants, air-cooled vs. water-cooled condensers) that may be operated via similar thermodynamic principles. This may be inter alia achieved by performing a model calibration, which may be considered as retraining the model based on training data that relates to different cooling system configurations. More specifically, the trained machine learning model may be retrained on an additional training dataset based on operational parameters obtained from other cooling processes / system types. Note that the retraining may not require samples from both classes (i.e., optimal and suboptimal) but may only be performed on samples that are classified as "optimal". For instance, a new cooling system may be set up and collect operational data for a period of time upon commissioning. It can be assumed that the system components are in good condition during this period of time and thus all samples can automatically be labeled as "optimal" or "efficient operation". This is advantageous since it removes the need for manual labeling and also the human bias that might come from human labeling. During retraining, a decision boundary or hyperplane of the trained machine learning model is adjusted to reflect the operational data from the new cooling system. Note that retraining may have several advantages over training a model from scratch such as less required resources, less required training data and faster training time.

[0094] In addition, the machine learning model may be further improved through the addition of historical data to the training set. For instance, longitudinal historic data about operational parameters during cooling cycles or energy efficiency predictions that were made by the model. Moreover, the model may be adapted to predict a time period until the next required maintenance.

[0095] Subsequently, the model is built, e.g., the trained model is stored for being used in conjunction with a cooling system.

[0096] The model running step may comprise the deployment of the trained machine learning model in the operating environment of the industrial cooling system to perform predictions on unseen data. The trained machine learning model might be deployed and stored on a local processor in a local area network (LAN) configuration in close proximity to the cooling equipment. This might be advantageous with regard to security concerns. A further option may be to deploy and store the trained machine learning model remotely, wherein the cooling system is operatively connected via an internet link to a remote processing service such as a "cloud service". The cloud service system may be accessed by the system operator via a personal login with credentials to reach information related to the cooling system. In addition, the operator may configure the system to send notifications (e.g., e-mail, text message) in case of an inefficient operating situation.

[0097] During the model export step, the model may be exported to be retrained using training data from various cooling systems. In a subsequent step, prediction results may be provided to a customer interface such that an operator of the system may review or obtain the prediction results.

[0098] Additionally, the operator may have the opportunity to configure the system to send notification such as an E-Mail alert or text message in case of a critical inefficient operating situation is detected. Such notification may also be sent directly to a trusted service provider. This may be advantageous since the service provider may directly take action to perform services such as ordering spare parts, scheduling maintenance and adjusting operating parameters including condenser discharge pressure optimization.

[0099] In case of a locally operated system, the notification may be transported via data exchange to a SCADA system and / or a dashboard solution. Accordingly, the machine learning model may be used in conjunction with a reporting functionality that prompts human interaction and further improves optimization of the cooling system efficiency.

[0100] Preferably, the trained machine learning model is one or more of a Long Short-Term Memory, LSTM, Network, a Recurrent Neural Network, RNN, a Vector Error Correction Model, VECM, a Support Vector Regression, SVR, model, a Physics-Informed Neural Network, PINN, a Deep Q-Networks, DQN, a Vector Auto Regression, VAR, model or an Autoencoder Neural Network.

[0101] It will be apparent that systems and / or methods described herein may be implemented in different forms of hardware, firmware, or a combination of hardware and software. The actual specialized control hardware or software code used to implement these systems and / or methods is not limiting of the aspects. Thus, the operation and behavior of the systems and / or methods were described herein without reference to specific software code-it being understood that software and hardware can be designed to implement the systems and / or methods based on the description herein.

[0102] Even though particular combinations of features are recited in the claims and / or disclosed in the specification, these combinations are not intended to limit the disclosure of various aspects. In fact, many of these features may be combined in ways not specifically recited in the claims and / or disclosed in the specification. Although each dependent claim listed below may directly depend on only one claim, the disclosure of various aspects includes each dependent claim in combination with every other claim in the claim set. A phrase referring to "at least one of" a list of items refers to any combination of those items, including single members. As an example, "at least one of: a, b, or c" is intended to cover a, b, c, a-b, a-c, b-c, and a-b-c, as well as any combination with multiples of the same element (e.g., a-a, a-a-a, a-a-b, a-a-c, a-b-b, a-c-c, b-b, b-b-b, b-b-c, c-c, and c-c-c or any other ordering of a, b, and c).

[0103] No element, act, or instruction used herein should be construed as critical or essential unless explicitly described as such. Also, as used herein, the articles "a" and "an" are intended to include one or more items, and may be used interchangeably with "one or more." Furthermore, as used herein, the terms "set" and "group" are intended to include one or more items (e.g., related items, unrelated items, a combination of related and unrelated items, and / or the like), and may be used interchangeably with "one or more." Where only one item is intended, the phrase "only one" or similar language is used. Also, as used herein, the terms "has," "have," "having," and / or the like are intended to be open-ended terms.

[0104] As used herein, the phrase "based on" shall not be construed as a reference to a closed set of information, one or more conditions, one or more factors, or the like. In other words, the phrase "based on A" (where "A" may be information, a condition, a factor, or the like) shall be construed as "based at least on A" unless specifically recited differently.

[0105] As used herein, the term "or" is an inclusive "or" unless limiting language is used relative to the alternatives listed. For example, reference to "X being based on A or B" shall be construed as including within its scope X being based on A, X being based on B, and X being based on A and B. In this regard, reference to "X being based on A or B" refers to "at least one of A or B" or "one or more of A or B" due to "or" being inclusive. Similarly, reference to "X being based on A, B, or C" shall be construed as including within its scope X being based on A, X being based on B, X being based on C, X being based on A and B, X being based on A and C, X being based on B and C, and X being based on A, B, and C. In this regard, reference to "X being based on A, B, or C" refers to "at least one of A, B, or C" or "one or more of A, B, or C" due to "or" being inclusive. As an example of limiting language, reference to "X being based on only one of A or B" shall be construed as including within its scope X being based on A as well as X being based on B, but not X being based on A and B.

[0106] Further, process diagrams such as FIG. 3 does not necessarily indicate a particular order or sequence of steps. For example, steps may also be performed in a different order or, if hardware capabilities allow it, simultaneously, without deviating from the scope of the present disclosure.

[0107] The foregoing disclosure provides illustration and description but is not intended to be exhaustive or to limit the aspects to the precise form disclosed. Modifications and variations may be made in light of the above disclosure or may be acquired from practice of the aspects. As used herein, the term component is intended to be broadly construed as hardware, firmware, or a combination of hardware and software. As used herein, a processor is implemented in hardware, firmware, or a combination of hardware and software.

Claims

1. A computer-implemented method for improving energy consumption of a cooling system comprising one or more condensers and one or more compressors, the method comprising: receiving previous sensor data associated with a cooling system; receiving a wet bulb temperature prediction; predicting a cooling system behaviour based on the previous sensor data and the wet bulb temperature prediction, using a trained machine learning model; and adjusting one or more cooling system parameters based on the prediction.

2. The computer-implemented method according to any one of the preceding claims, wherein the wet bulb temperature prediction is derived from the weather forecast, preferably for the upcoming hour.

3. The computer-implemented method according to any one of the preceding claims, wherein predicting the cooling system behaviour comprises: predicting a plurality of cooling system behaviours based on the previous sensor data and the wet bulb temperature prediction, wherein each predicted cooling system behaviour of the plurality of predicted cooling system behaviours is associated with a corresponding cooling system parameter adjustment.

4. The computer-implemented method according to the preceding claim, further comprising: selecting the predicted cooling system behaviour having a lowest predicted power consumption from the plurality of predicted cooling system behaviours; and adjusting the corresponding cooling system parameter accordingly.

5. The computer-implemented method according to the preceding claim, further comprising: determining a baseline power consumption of the cooling system, wherein selecting the predicted cooling system behaviour includes comparing the selected predicted cooling system behaviour with the baseline power consumption.

6. The computer-implemented method according to any one of the preceding claims, wherein adjusting the one or more cooling system parameters further comprises: determining a boundary condition based on one or more of a predefined domain rule, a data-driven threshold and / or a safety margin; and adjusting the one or more cooling system parameters based on the boundary condition.

7. The computer-implemented method according to any one of the preceding claims, wherein the predicted cooling system behaviour includes a one or more of an expected fluid pressure, an expected condensing temperature, and / or an expected total power consumption value of the cooling system.

8. The computer-implemented method according to any one of the preceding claims, wherein the previous sensor data comprises one or more of a measured condensing temperature, a compressor cooling capacity, a measured compressor driving power, a measured condenser fan speed, a condenser capacity, a saturated discharge temperature, a saturated suction temperature, an ambient air temperature, a condenser inlet temperature and / or a condenser outlet temperature.

9. The computer-implemented method according to any one of the preceding claims, further comprising: aggregating a plurality of previous sensor data points into one or more aggregated previous sensor data points, wherein the cooling system behaviour is predicted based on the aggregated previous sensor data and the wet bulb temperature prediction, wherein preferably each type of previous sensor data is aggregated into a corresponding aggregated previous sensor data point.

10. The computer-implemented method according to any one of the preceding claims, wherein the cooling system parameter includes one or more of a compressor power supply, a compressor cooling capacity, a set condensing temperature, a condenser fan speed, an expansion valve setting, a compressor driving power, a condenser capacity, a suction temperature, a discharge pressure, a discharge temperature or combinations thereof.

11. A cooling system comprising: one or more condensers; one or more compressors; and means for carrying the computer-implemented method according to any one of the preceding claims.

12. A computer-implemented method for training a machine learning model configured to predict a cooling system behaviour of a cooling system comprising one or more condensers and one or more compressors, the method comprising: obtaining training data from one or more cooling systems comprising a plurality of parameter sets comprising one or more of: a discharge temperature from the compressor, a discharge pressure from the compressor, a suction pressure from the compressor, a suction temperature from the compressor, a compressor cooling capacity; a power consumption of the compressor, a power consumption of the condenser, a fan speed of the condenser, a condenser capacity; a fan power consumption of the condenser, a temperature at the condenser input, a temperature at the condenser output, an expansion valve position, a wet bulb temperature, an ambient air temperature, a cooling load demand, a time of day, seasonal pattern; training the machine learning model based on the training data to generate a trained machine learning model.

13. A trained machine learning model configured to predict a cooling system behaviour of a cooling system comprising one or more condensers and one or more compressors, wherein the trained machine learning model is trained according to the method of the preceding claim.

14. The trained machine learning model of the preceding claim, wherein the machine learning model is one or more of a Long Short-Term Memory, LSTM, Network, a Recurrent Neural Network, RNN, a Vector Error Correction Model, VECM, a Support Vector Regression, SVR, model, a Physics-Informed Neural Network, PINN, a Deep Q-Networks, DQN, or a Vector Auto Regression, VAR, model.

15. A computer program comprising instructions which, when executed by a processor, cause the processor to carry out the computer-implemented method according to any one of claims 1 to 10.

Images