Analysis of Time Series of Physical Characteristics in an Ice Rink Using IoT and Machine Learning
Summary
In the face of the climate crisis and rising energy costs, improving energy efficiency is increasingly essential. Nations seek to reduce CO₂ emissions, and businesses aim to optimise energy expenditure. Digitalisation and the adoption of Internet of Things (IoT) technologies create opportunities for applying artificial intelligence (AI) to energy consumption management.
This article reviews AI technologies applied to energy management, using an ice rink as a case study. It analyses real consumption data collected from an operational site and examines neural network time-series modelling methods for forecasting and optimising control. The modelling results demonstrate the potential of predictive algorithms to reduce energy costs and improve operational efficiency in ice arenas.
Introduction
Modern economic and environmental realities force enterprises and governments to reassess how energy is consumed. Research shows that electricity expenses for medium and large Russian companies can account for 6%–20% of total production and service costs.
Sports facilities — especially ice rinks — have high energy consumption because they must maintain specific climate conditions over large indoor spaces with complex equipment. Efficient energy management on such sites requires balancing service quality and consumption optimisation. An ice arena’s numerous zones and activities make predictive energy and comfort forecasting challenging.
The growth of IoT and digitalisation makes it possible to collect large real-time datasets, providing the foundation for applying AI and machine learning (ML) methods to energy management.
The research presented here explores predictive analytics and deep learning models for forecasting energy management parameters in ice arenas, analysing real-world data and modelling results based on neural network architectures.
Relevance of the Research
Global warming and related climate change demand urgent action to cut greenhouse gas emissions. Enhancing energy efficiency is one of the most effective measures to achieve this — it not only lowers CO₂ emissions but also reduces operational costs.
In the sports industry, energy constitutes a major portion of operating expenses. Ice rinks, in particular, consume significant energy to maintain ice surfaces, HVAC systems, lighting, and support systems. Optimising consumption can lead to substantial savings and a reduced environmental footprint.
Overview of Approaches
Existing strategies for optimising energy management in sports facilities can be grouped into two main categories: physics-based models and data-driven methods.
Physics-based methods build mathematical representations of physical processes. For example, one study developed a physical model to control thermal comfort in a hot, humid football stadium using computational fluid dynamics. Another applied a genetic optimisation algorithm to reduce HVAC energy use in a swimming pool.
While these models can accurately describe systems from first principles, they require detailed building information and are complex to develop for large structures.
Machine learning methods, enabled by IoT data volumes, use historical datasets to train models capable of forecasting system behaviour and optimising control. Forecast-based control, such as Model Predictive Control (MPC), forecasts future system states and determines optimal inputs within constraints.
Recurrent neural networks (RNNs) — especially LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) — are effective for time-series forecasting because they learn long-term dependencies in noisy, sequential data. Convolutional neural networks (CNNs) are also applied for feature extraction in time series. Hybrid models combining RNNs and CNNs leverage both approaches.
When applied to ice rinks, such methods can:
forecast energy and temperature dynamics;
optimise equipment operation to cut energy costs without compromising ice quality;
account for external and internal influences (weather, event schedules, attendance);
respond in real time when integrated with IoT systems.
Dataset Description
For this study, IoT-collected data from a real ice arena was used. The dataset includes one-minute measurements of various parameters, enabling high-resolution temporal analysis:
temp_ice — ice surface temperature (°C)
consumption_compressors — total compressor energy consumption (kWh)
temp_outside — outside air temperature (°C)
hum_outside — outside relative humidity (%)
temp_inside — interior air temperature (°C)
hum_inside — interior relative humidity (%)
max_motion — movement/activity level on the ice
illuminance — light intensity
CO₂ — carbon dioxide level (ppm)
temp_condensation — refrigerant condensation temperature (°C)
temp_supply_glyc — glycol supply temperature (°C)
temp_return_glyc — glycol return temperature (°C)
Before modelling, anomalies and missing values were cleaned to improve data quality and model accuracy.
Ice surface temperature — the key target variable — often deviated from the ideal –3.5 °C, leading to increased energy use and lower ice quality.
Internal: compressor and pump performance, refrigerant flow and temperatures
Methodology
To predict ice surface temperature, an LSTM-based recurrent neural network model was developed. It used the previous 60 minutes of parameters to forecast the next 30 minutes. LSTM was chosen because of its ability to capture long-term temporal dependencies and handle noisy sequences.
The dataset was split into:
70% training
15% validation
15% testing
Regularisation techniques like dropout and early stopping were used to prevent overfitting. Training was performed using the Adam optimiser and mean squared error (MSE) loss.
Discussion of Results
The LSTM model achieved the following forecasting performance:
MSE: 0.25 °C²
RMSE: 0.5 °C
MAE: 0.4 °C
This indicates that the model captured the overall ice temperature dynamics. However, it struggled with rapid temperature spikes caused by water resurfacing, reducing its effectiveness for real-time compressor control.
Most of the missed resurfacing events suggest limitations in the model’s ability to capture short-term anomalies.
Future Directions
To improve forecasting accuracy for sudden events, more advanced approaches may be needed. Autoencoders can encode complex dependencies and help reconstruct missing or anomalous points. Variational autoencoders (VAEs) could generate additional features or fill gaps, enhancing prediction.
Transformers — proven effective in sequential data processing — may outperform traditional RNNs, handling long-range dependencies and predicting complex events like ice resurfacing.
Incorporating these methods could significantly enhance forecast precision and the system’s adaptability to diverse operational scenarios.
Conclusion
Applying machine learning to energy management in ice rinks shows considerable promise. The LSTM-based model performed well overall, supporting its use in refrigeration control systems. However, limitations with abrupt changes highlight the need for further model refinement.
Future research into autoencoders and transformer-based architectures could enhance model performance, reducing energy consumption, improving ice quality, and extending equipment life.
Overall, these advancements will contribute to greater economic efficiency and environmental sustainability for sports facilities — an important goal for both businesses and society.