The increasingly inordinate production of productivity-enhancing technologies in today's society has led to an increased risk of damaging the environment from which key resources are derived, if conducted in an uncontrolled manner. It is therefore necessary to optimise strategies to make technological progress more sustainable. Artificial Intelligence (AI) plays a crucial role in this scenario and it is therefore crucial to consider sustainable AI as an integral step in the whole process. In this work we present the results and the topics under investigations in our laboratory focused on the responsible use of AI to meet technological improvement.
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License based technologies makes it necessary to align the poAttribution 4.0 International (CC BY 4.0).
but they all have in common the idea of creating machines that can think like humans, with the intention of improving the efectiveness and eficiency of their activities and thus ultimately the quality of life [1].
Conversely, however, inordinate expansion can have a negative impact on the environment, and there is no ucts (or innovation) requires attention to its impact on the environment. For instance, mass farming has been shown to impact biodiversity, mass production of clothing has an impact on the world’s water reserves, and e-waste releases chemicals and poisons into the water and soil where it is dumped. lution of technology should be geared towards the wise choice of resources to reduce environmental impact. In these terms, sustainable Artificial Intelligence (AI) can help optimising productions according to the specific tasks required, developing an AI that is compatible with the preservation of environmental resources for current and future generations, through the economic models of societies, and with the fundamental social values of a given society [2].
nized by CINI, May 29–31, 2023, Pisa, Italy ∗Corresponding author.
Intelligent Transportation Systems (ITSs) have been deincreasing urbanization and the latest advances in technology have made it a timely and extremely relevant topic [3, 4]. ITSs aim to improve transport systems in all aspects and concern the design, analysis, and control of information technology by integrating data from diferent sources (e.g., GPS sensors, cameras, LIDAR sysdirections: developing an open-source toolbox for mapmatching and road pothole recognition and segmentation. The two projects can be combined to create routing software that considers road quality, enabling drivers to choose smoother routes. This could enhance driver safety and reduce fuel consumption. 2.1. PyTrack The exponential growth of IoT devices, smartphones, smartwatches, and vehicles equipped with positioning technology, such as Global Positioning System (GPS) modules, has boosted the development of location-based services for several applications in Intelligent Transportation Systems. However, the inherent error of locationdoubt that any global and pervasive production of prod- veloped since the second half of the 20th century. Both
It is therefore necessary to bear in mind that any evo- tems and so on) [3]. We serve this paradigm under two sitioning trajectories to the actual underlying road network, a process known as map-matching [5]. To the best of our knowledge, there are no comprehensive tools that allow us to model street networks, conduct topological and spatial analyses of the underlying street graph, perform map-matching processes on GPS point trajectories, and deeply analyze and elaborate these reconstructed trajectories. To address this issue, we present PyTrack, an open-source map-matching-based Python toolbox designed for academics, researchers and practitioners that integrate the recorded GPS coordinates with data provided by the OpenStreetMap, an open-source geographic information system.
With reference to the figure 1 in a nutshell, the algorithm performs the following steps: • Using the API exposed by Google Map, the algorithm matches each coordinate point provided by the user’s GPS as input to the closest corresponding point in a navigable route on the map; • Finds the best trajectory that links all the matched points and maintain its coherence with the street avoiding one-way streets and roundabouts in the wrong direction; • Using the API made available by OpenStreetMap, it performs a compact oversampling of the entire space between all pairs of snapped coordinates following the best trajectory and thus generates a video with the corresponding sequence of frames oriented in the same direction as the driver would be oriented if driving.
experimental setup for capturing data and simulating a smart vehicle. To this end, a stereo camera was identified For more information on PyTrack, users are encouraged that could capture multimodal information, including a to visit the oficial repository 1. depth map for observing the road surface. To complete this work, the next steps involve gathering 2.2. Road Quality Evaluation and annotating a dataset through an acquisition phase and validating the model’s performance using computer We aim to improve ITS also by implementing a system vision techniques to detect potholes. A representation of to perform road condition analysis using computer vi- the described pipeline is depicted in 2. sion techniques applied directly to information extracted from an onboard vehicle camera. This can serve ITSs with innovative functionalities to facilitate and improve 3. The shift towards the the transportation sector with a positive return on com- Renewable Energy Sources: a munity welfare. This technology could be applied to recognize and locate road irregularities to provide a high- photovoltaic case study level autonomous driving system with information to avoid potholes.
The current work has focused on exploring the latest computer vision methods for detecting potholes and road cracks. Specifically, the emphasis has been on using multimodal techniques, which involve using heterogeneous data sources to analyze the problem from various points of view and thus improve the model’s overall performance. In addition, the focus was to find an appropriate
reduce greenhouse gas emissions, and macroeconomic and geopolitical instability, the integration of renewable energy into modern power grids is steadily increasing.
The shift towards renewable energy sources is crucial to a sustainable, afordable, accessible, clean, and low-carbon future, reducing polluting emissions and dependence on fossil fuels.
In this regard, photovoltaic (PV) energy is rapidly emerging as one of the world’s most promising renewable energy sources, playing a crucial role in accomplishing lfexibility and, those that rely on statistical approaches, various climate protection goals [6, 7, 8], indeed, thanks even if they ofer a good compromise between accuracy to the guarantees of low carbon consumption, adaptabil- and model simplicity, may be unsuitable for discovering ity to diferent applications and the advantage of keeping non-linear relationships and are highly dependent on installation and maintenance costs low, ascribing it as a the quality of the available data. On the other side, due sustainable energy source [9]. Compared to fossil fuel- to their ability to discover complex relationships, deal derived energy, green energies are substantially more with unstructured data, and superior performance, AIsustainable. However, their inherent intermittent nature based models have focused the research in recent years. does not guarantee constant production flows, causing Among these methods, recurrent neural networks (RNNs) imbalances in electrical grids that ultimately limit large- and convolutional neural networks (CNNs) are the most scale adoption [10]. Nevertheless, in recent years, with used DL-based architecture providing state-of-the-art the rapid development of IoT technology, largely powered performance in PV power forecasting [14]. by artificial intelligence, various smart applications have In photovoltaic systems, the amount of energy probeen applied to many fields, and a forecasting system duced is heavily influenced by weather factors, such as sothat simulates hourly global solar irradiance predictions lar radiation. While deep learning methods have shown has also been proposed [11, 12]. promising results in forecasting PV power production,
Photovoltaic energy generation forecasting could in they ignore the underlying physical prior knowledge of fact help resolve these imbalances and uncertainties, facil- the phenomenon. This is where NWP-based methods itating the introduction of renewable energy sources into can be incredibly advantageous, as they provide valumodern power grids [13, 14, 15]. Accurate forecasting of able insights about the weather factors that afect energy photovoltaic production emerges therefore as an essential production and, therefore, can help improve forecasting stakeholder to realise the full potential of PV systems and accuracy. Despite the potential benefits, relatively few provide grid operators and energy traders with valuable works have explored the combination of NWP and DL insights and decision-making information to optimise models [17]. maintenance strategies, plan the development of new In this respect we propose MATNet, a novel selfplants, mitigate operational and management challenges, attention-based architecture for multi-step day-ahead and improve economic returns on investment [16]. photovoltaic power production forecasting, combining
For this reason, several methods for forecasting photo- the advantages of a deep learning approach with the a privoltaic energy production have been developed recently, ori knowledge of the phenomenon provided by physicaldivisible in physics-based and data-based, in turn the based models. latter divisible into statistics-based and AI-based. The contributions of this work are summarised as fol
While methods based on physical data, also known lows: as Numerical Weather Prediction (NWP), can be used to emulate complex systems, they can often reveal a lack of • We propose a novel self-attention transformerbased architecture for multivariate multi-step day-ahead photovoltaic power production fore- approach to reorganise the entire agricultural system casting. The attention mechanism is a vital part towards sustainable low-input, high-eficiency agriculof the architecture enabling the model to focus ture. This new approach mainly benefits from the emeron input data elements dynamically. gence and convergence of diferent technologies, includ• The proposed architecture consists of a hybrid ap- ing miniaturised computing components, automatic conproach that combines the ability to generalise and trol, field and remote sensing, and mobile computing. automatically capture complex patterns and rela- The agricultural industry is therefore now able to coltionships in the data, typical of the deep learning lect more comprehensive data on production variability paradigm, with the prior physical knowledge of over time and considering diferent geographical areas photovoltaic power generation of NWP methods. of interest [18, 19]. The next step for the PA is to be able By combining these two approaches, our model to respond to this variability of optimisation demands can achieve a more accurate and robust PV power on a finer scale and thus inherent to, for example, indigeneration forecast. vidual crops of specific products, trying to apply mass • We fed the proposed model with historical photo- knowledge and adapt it to detailed cases. voltaic data and historical and forecast weather In these terms, an increasingly popular practice for data. The historical and forecast weather data are such a task is the use of AI, as a useful and efective conceptually diferent as they observe the phe- tool for maximising production yield, minimising waste nomenon at diferent points in time. Therefore, and energy consumption. Indeed, AI in PA can involve the architecture handles these input branches better utilization of a farm’s resources, such as fertilthrough joint fusion performed at diferent levels izers, herbicides, irrigation, and seeds allowing for the of abstraction. correct intervention at the right place and time. PA aims • We propose a dense interpolation module to to improve crop production by maximizing yields with simplify the high-dimensional representation re- minimal chemical application, where, as example, a field turned by the attention-based module. can be divided into zones, each receiving a customized • We evaluate the model’s efectiveness by compar- amount of resources based on diferent landscape types ing it extensively with the Ausgrid benchmark and management history [20]. dataset using diferent regression performance metrics, ablation studies, and data perturbations. 4.1. Low Orchard Productivity The results show that the proposed MATNet ar- Assessment chitecture with an RMSE score equal to 0.0673 significantly outperforms the current state-of-theart methods. On the other hand, the ablation study of the diferent modalities study highlights the crucial role that weather forecasts play in the overall performance of our model.
the field of fruit and vegetable production. Indeed, in partnership with Zespri Kiwifruit, we are developing an AI-based application to map the Kiwifruit Vine Decline Syndrome (KVDS) phenomenon on G3 using satellite image data. Measuring the impact of KVDS on Yellow Kiwi plants is useful for assessing the expected decrease 4. Precision Agriculture in the year’s production, the quality of the fruit that will be harvested, and the prospects of production and proBefore the advent of agricultural mechanisation, the very ductivity from a historical point of view. To this end, we small size of plots allowed farmers to vary treatments retrieve from the Sentinel-2 satellites all the historical manually. However, with the enlargement of fields and imaging data available for the partner farms of Zespri for intensive mechanisation, it has become increasingly dif- the G3 project. The Sentinel-2 satellites provide 13 specifcult to take into account the variability within the field tral channels in the visible/near-infrared and short-wave without a revolutionary development of technologies, infrared spectral range with 10 meters spatial resolution. where crops must face several challenges arising from Starting from the spectral channels, we are computing sub-optimal management of agricultural resources and well-known vegetation indices such as NDVI, NDMI and increased pressure from biotic and abiotic stress factors, NDRE [21] to map the vegetation health of considered and therefore the need for a major restructuring of re- farms during time. Figure 4 shows a schematic of the sources has increased, while also seeking solutions with developed pipeline for NDVI computation. Note that a reduced impact on the entire ecosystem as much as pos- for NDVI estimation, we need from Sentinel-2 the RED sible. The proposed pipeline and consequent forecasting and NIR spectral bands. Merging all the retrieved data are depicted in Figure 4. points allows for estimating a multivariate time series of
In the recent years, Precision Agriculture (PA) has the considered indexes performing inter- and intra-farm become widespread and conceptualised as a systemic statistical analysis of both. In the former case, we are considering all the farms provided by Zespri to compute a global stress indicator across the years. This analysis allows aggregating farms into stress areas giving an insight about the actual condition of KVDS. In the latter case, we aim to find local stress regions within each farm, exploiting classical AI algorithms like clustering.
Finally, to visualize the obtained results, a dashboard will be developed.
As a future development, we plan to employ Generative Models to design an image-to-image translation framework to remove the presence of hail nets from the considered crops which reduces the reliability of the calculated indexes. Another possible direction relies on developing an AI pipeline to detect clouds from downloaded images to understand if they overlap or not the considered farm. This allows for increasing the amount of available data from satellites.
produced the double efect of promoting productivity in various fields of daily life together with the risk of damaging the environment due to the uncontrolled exploitation of resources, thus requiring more sustainable progress. For our side, we have therefore presented three topics under investigation in our laboratory, namely PyTrack, an open source Python toolbox which aims to reconstruct the best trajectories starting from GPS coordinates and subsequently combine them with the images produced by the OpenStreetMap API to allow an assessment of the quality of the road ruined by potholes and cracks. As a second contribution we propose MATNet, a novel self-attention-based architecture for multivariate prediction of day-ahead photovoltaic energy production. Its evaluation against reference datasets far exceeds current state-of-the-art methods. Finally, in the field of precision agriculture, we presented a collaboration with Zespri Kiwifruit through an AI-based application with the aim of mapping the phenomenon of kiwi vine wasting syndrome over time using satellite image data.
newable and Sustainable Energy Reviews 13 (2009) 2096–2103.
We acknowledge FS Technology, Zespri Kiwifruit, Ing. [10] R. Meenal, D. Binu, K. Ramya, P. A. Michael, Francesco Conte, PhD, Unit of Innovation, Entrepreneur- K. Vinoth Kumar, E. Rajasekaran, B. Sangeetha, ship Sustainability, Department of Engineering Univer- Weather forecasting for renewable energy system: sity Campus Bio-Medico of Rome. We also acknowledge a review, Archives of Computational Methods in ifnancial support from PNRR MUR project PE0000013- Engineering 29 (2022) 2875–2891. FAIR and from the project n. F/130096/01-05/X38 - Fondo [11] P. F. Jiménez-Pérez, L. Mora-López, Modeling and per la Crescita Sostenibile - ACCORDI PER L’INNO- forecasting hourly global solar radiation using clusVAZIONE DI CUI AL D.M. 24 MAGGIO 2017 - Ministero tering and classification techniques, Solar Energy dello Sviluppo Economico (Italy), iii) Programma Opera- 135 (2016) 682–691. tivo Nazionale (PON) “Ricerca e Innovazione” 2014-2020 [12] H.-T. Yang, C.-M. Huang, Y.-C. Huang, Y.-S. Pai, CCI2014IT16M2OP005 Azione IV.4. A weather-based hybrid method for 1-day ahead hourly forecasting of pv power output, IEEE transReferences actions on sustainable energy 5 (2014) 917–926. [13] P. Gupta, R. Singh, Pv power forecasting based on [1] M. A. Goralski, T. K. Tan, Artificial intelligence and data-driven models: a review, International Journal sustainable development, The International Journal of Sustainable Engineering 14 (2021) 1733–1755. of Management Education 18 (2020) 100330. [14] S. Aslam, H. Herodotou, S. M. Mohsin, N. Javaid, [2] A. Van Wynsberghe, Sustainable ai: Ai for sustain- N. Ashraf, S. Aslam, A survey on deep learning ability and the sustainability of ai, AI and Ethics 1 methods for power load and renewable energy fore(2021) 213–218. casting in smart microgrids, Renewable and Sus[3] L. Zhu, F. R. Yu, Y. Wang, B. Ning, T. Tang, Big tainable Energy Reviews 144 (2021) 110992. data analytics in intelligent transportation sys- [15] D. Kaur, S. N. Islam, M. A. Mahmud, M. E. Haque, tems: A survey, IEEE Transactions on Intelli- Z. Y. Dong, Energy forecasting in smart grid sysgent Transportation Systems 20 (2019) 383–398. tems: recent advancements in probabilistic deep URL: https://ieeexplore.ieee.org/document/8344848. learning, IET Generation, Transmission & Distridoi:10.1109/TITS.2018.2815678. bution 16 (2022) 4461–4479. [4] A. Haydari, Y. Yılmaz, Deep reinforcement [16] G. Alkhayat, R. Mehmood, A review and taxonomy learning for intelligent transportation systems: of wind and solar energy forecasting methods based A survey, IEEE Transactions on Intelligent on deep learning, Energy and AI 4 (2021) 100060. Transportation Systems 23 (2022) 11–32. URL: [17] M. Bai, Y. Chen, X. Zhao, J. Liu, D. Yu, Deep attenhttps://ieeexplore.ieee.org/document/9146378. tion convlstm-based adaptive fusion of clear-sky doi:10.1109/TITS.2020.3008612. physical prior knowledge and multivariable histor[5] Z. Huang, S. Qiao, N. Han, C.-a. Yuan, X. Song, ical information for probabilistic prediction of phoY. Xiao, Survey on vehicle map matching tovoltaic power, Expert Systems with Applications techniques, CAAI Transactions on Intelligence 202 (2022) 117335.
Technology 6 (2021) 55–71. URL: https://ietresearch. [18] N. Zhang, M. Wang, N. Wang, Precision agriculonlinelibrary.wiley.com/doi/abs/10.1049/cit2.12030. ture—a worldwide overview, Computers and elecdoi:10.1049/cit2.12030. tronics in agriculture 36 (2002) 113–132. [6] International Renewable Energy Agency, So- [19] A. Sharma, A. Jain, P. Gupta, V. Chowdary, Malar energy, 2020. URL: https://www.irena.org/ chine learning applications for precision agriculEnergy-Transition/Technology/Solar-energy, [On- ture: A comprehensive review, IEEE Access 9 (2020) line; accessed 2023-02-20]. 4843–4873. [7] International Energy Agency, Solar-PV, 2022. URL: [20] J. D. Rudd, G. T. Roberson, J. J. Classen, Applihttps://www.iea.org/reports/solar-pv, [Online; ac- cation of satellite, unmanned aircraft system, and cessed 2023-02-20]. ground-based sensor data for precision agriculture: [8] Y. Zhou, Artificial intelligence in renewable sys- A review, in: 2017 ASABE Annual International tems for transformation towards intelligent build- Meeting, American Society of Agricultural and Bioings, Energy and AI (2022) 100182. logical Engineers, 2017, p. 1. [9] P. Nema, R. Nema, S. Rangnekar, A current and [21] J. Mezera, V. Lukas, I. Horniaček, V. Smutnỳ, J. Elbl, future state of art development of hybrid energy Comparison of proximal and remote sensing for system using wind and pv-solar: A review, Re- the diagnosis of crop status in site-specific crop management, Sensors 22 (2021) 19.