A transformer-based architecture: The informer model for uav power consumption estimation

Abstract

Unmanned Aerial Vehicles (UAVs), commonly known as drones, are increasingly used in applications ranging from package delivery to surveillance. One critical challenge limiting wider adoption of UAVs is their energy constraint, small onboard batteries severely restrict flight duration and range. UAV power consumption is influenced by numerous dynamic factors, including weather conditions, payload weight, flight speed, and trajectory path. These environmental and operational uncertainties lead to significant variability in energy usage, underscoring the need for reliable power consumption prediction models. Accurate forecasts of a drone's power requirements are essential to ensure safe operations, prevent mid-flight battery depletion, and optimize mission planning. Studies highlight that improving prediction accuracy can help extend flight times, manage payloads effectively, and implement contingency measures to enhance overall UAV reliability. In light of these needs, recent research has turned to advanced machine learning techniques for modelling UAV energy consumption. Integrating deep learning with UAV energy management allows data-driven prediction of power usage, which can optimize resource allocation and improve system performance. Power consumption forecasting has emerged as a particularly crucial area of study, as it directly affects UAV operational efficiency. This thesis addresses the challenge by exploring a state-of-the-art transformer-based architecture, known as the Informer model, to estimate UAV power consumption. The Informer model is a recent development in time-series forecasting that is designed to handle long sequences efficiently. By applying this model to a comprehensive UAV flight dataset, the study aims to capture the complex relationships between flight conditions and power use, ultimately improving the accuracy of power consumption predictions for drones. A key trend in recent literature is the introduction of attention mechanisms and Transformer-based models to UAV energy prediction. Attention allows models to weight the influence of different time steps or features, which is valuable given the multivariate nature of drone flight data (speed, altitude, wind, etc.). In fact, studies have shown that adding self-attention to sequence models (such as augmenting GRUs with self-attention) yields noticeably better performance in power consumption forecasting. By focusing on critical input features during a flight, attention-based models can more accurately predict power usage fluctuations caused by factors like sudden wind gusts or changes in payload. These promising results set the stage for applying Transformer architectures (originally developed for natural language processing) to UAV time-series data. The Transformer model's capacity to capture long-range dependencies without recency bias makes it attractive for this domain. However, a vanilla Transformer can be computationally expensive for long flight sequences due to its quadratic complexity in sequence length. To overcome this, researchers developed the Informer architecture, which specifically targets long sequence forecasting with improved efficiency. Early evaluations of the Informer on benchmark datasets showed that it significantly outperforms previous forecasting methods while handling long sequences with reduced computational cost. These developments in the literature provide a strong foundation and motivation for the thesis to implement the Informer model for UAV power consumption estimation. The thesis adopts the Informer model, a Transformer-based neural network architecture, to perform multivariate time-series forecasting of UAV power consumption, by using a dataset which is a publicly available flight log repository collected via real time experiments. The choice of Informer is motivated by its ability to handle long sequences of data more efficiently than a standard Transformer. In essence, Informer is an encoder–decoder model tailored for long-term forecasting. the encoder processes a lengthy input sequence of historical data, and the decoder produces predictions for a future sequence (the target horizon). For UAV power estimation, this means the model takes a window of past flight telemetry (multiple time steps of sensor readings and power usage) and predicts the drone's power consumption over upcoming time steps. In applying the Informer model to the UAV dataset, the thesis follows a typical model development and evaluation pipeline. The cleaned and preprocessed data (with selected features) are divided into training and testing subsets, ensuring that the model learns from one portion of the flights and is validated on separate flights it has not seen. The multivariate time series from the training set are used to fit the Informer's parameters; this involves adjusting the model's internal weights so that its output power consumption sequence closely matches the actual recorded power usage. he training optimization minimizes error metrics such as Mean Squared Error (MSE) between the predicted and actual power values. Importantly, the evaluation on the test set provides an unbiased assessment of how well the model generalizes to new flight conditions. Performance is measured with standard forecasting accuracy metric, mean squared error. This research culminates in a better understanding of how modern deep learning architectures can be harnessed for UAV energy management. The thesis showed that a transformer-based model (Informer) can successfully estimate a drone's power consumption in flight with a high degree of accuracy. Through comprehensive experiments on a rich UAV flight dataset, the Informer model learned complex relationships between flight parameters and power usage that would be difficult to capture with conventional methods. The result is a forecasting system capable of predicting future power requirements over the course of a flight, enabling more informed decision-making for drone operations. For instance, with reliable power consumption predictions, operators or autonomous flight planners can schedule battery swaps or recharges proactively, adjust flight paths or speeds to conserve energy, and ensure that drones maintain sufficient power reserves for safe return-to-base. In conclusion, the thesis contributes an important case study of applying the Informer model to a practical engineering problem in the UAV domain. It demonstrates that multivariate time-series forecasting using transformers is not only feasible but advantageous for predicting drone energy usage. The work provides a foundation for future enhancements, such as integrating real-time prediction into UAV control systems or extending the model to include additional factors (like battery health or external weather forecasts). By improving the accuracy of power consumption estimation, this research helps pave the way for safer and more efficient UAV operations, allowing drones to fly smarter and make better use of their limited energy reserves.