Generative AI has rapidly progressed from chatbots and assistants to agents across diverse applications and domains. Generative AI agents demonstrate sophisticated operation through autonomy, tool use and decision making with minimal human input. Despite these performance gains, agents are still impacted by the foundational limitations of Generative AI models. Among these, hallucinations are a major limitation that affects agent operation in real-world settings, leading to risk and loss. Several recent work aim to address hallucinations through methods such as retrieval-augmented generation and reflection prompting, however, these only provide partial improvements. An effective yet underexplored approach is in the observability data generated by an agent in its deployed and operational settings. Drawing on agent observability data, this paper proposes a dual memory knowledge graph approach that integrates Semantic and Observability Memory to address hallucinations in Generative AI agents. Semantic Memory provides organized domain knowledge for precise factual grounding. Observability Memory transforms logs, traces, and execution results into agent validated planning histories. Hallucinations are then addressed by grounded planning in verified past interactions with known, reliable outcomes. This approach is evaluated in a two-stage experimental setup aligned with its dual memory design. Observability memory is evaluated on the HotpotQA dataset to assess its impact on reasoning grounding, using metrics that capture both factual accuracy and reasoning hallucinations. The SM3-Text-to-Query benchmark and Synthea-based medical QA datasets are used to assess factual grounding of the semantic memory. Results from both experiments demonstrate reductions in hallucinations, with semantic memory for contextual grounding reducing factual hallucinations, and observability memory for reasoning grounding reducing faithfulness hallucinations.

Solar photovoltaics are on track to becoming the largest renewable energy source by 2029. This means a rapid increase in the number of solar energy generation installations from residential roof-top systems to utility-scale power plants. The current industrial approaches towards predictive maintenance will be insufficient to manage and maintain the increasing numbers of such installations at peak performance. In this paper, we propose hyper-predictive maintenance as a novel approach based on Generative Artificial Intelligence (AI) agents for highly autonomous management of solar energy infrastructure. The proposed Agentic AI framework deploys multiple agents for baseline generation from solar installations, predictive model development, degradation estimation, degradation evaluation and predictive maintenance that combines baseline performance with contextual information to predict faults and potential causes. This framework is empirically evaluated in the real-world solar energy systems of a multi-campus tertiary education institution. The results of these experiments confirm the robust and accelerated performance of Generative AI agents for the hyper predictive maintenance of large-scale solar energy installations.

The rapid advancement of the energy sector in terms of diverse energy generation options and increasing energy consumption loads has eventuated the need for highly accurate demand forecasting methods. The prevalence of large volumes of energy data streams and sophisticated Artificial Intelligence (AI) algorithms has enabled a rapid transition to AI-based forecasting methods that are more accurate and computationally efficient. Despite this transition, demand forecasting during peak events and peak temporal periods continues to be a challenge due to the irregularity and transience of such events. Besides the challenge of managing supply and demand, the financial viability of forecasting is also questioned when the forecast decreases in accuracy during peak periods when the energy price is an increasing function. In this paper, we have set out to address the challenge of peak demand forecasting by specifically transforming both input vectors and input attributes of the smart meter data streams. Input vectors are transformed using active thresholding while input attributes are transformed into a feature subset using model explainability. We have evaluated the effectiveness of this data transformation on the current state-of-the-art AI for energy demand forecasting, gradient boosting. We conduct a comparative evaluation using two real-world energy consumption datasets drawn from the La Trobe Energy AI/Analytics Platform (LEAP), of La Trobe University’s Net Zero Carbon Emissions Program. The proposed approach surpasses the baseline approach in both datasets, with an improvement of 27% for the second dataset which is a high energy consumption setting.

Microgrids and energy platforms have become increasingly intricate in nature. This is especially true in platforms adopted in large multi functional complexes which are spread across geographies of varying climactic conditions. Many of these complexes have evolved over time as organisations have grown and diversified, this has resulted in a mix of infrastructure, technology, networks, systems and equipment woven in an inextricable web. Often managing such complex hierarchies involve interacting with heterogeneous management platforms from various vendors and built on different platforms each providing the solution to a piece of the ’energy puzzle’ of the organisation. The complexities involved in managing such a network of systems are manifold and require extensive resourcing and expertise. Sustainability and Net Zero carbon emission initiatives that aim to achieve international and national targets add a further layer of complexity to the task at hand. In this paper, we propose a Cloud Edge architecture that leverages Artificial Intelligence (AI) and data analytics for microgrid energy optimisation and net zero carbon emissions. This architecture provides an intelligent and cohesive abstraction to assist in cataloging, unifying and managing the complexities of microgrids and enabling sustainable management of energy. The proposed architecture has been operationalised as the energy management and optimisation platform at a multi-campus, multi-functional tertiary education institution. Empirical evaluations conducted on this deployment have generated results that confirm the function and effectiveness of this architecture in addressing the emerging and evolving challenges of microgrid energy optimisation and net zero carbon emissions.

Hyperdimensional (HD) Computing leverages random high dimensional vectors (>10000 dimensions) known as hypervectors for data representation. This high dimensional feature representation is inherently redundant which results in increased robustness against noise and it also enables the use of a computationally simple operations for all vector functions. These two properties of hypervectors have led to energy efficient and fast learning capabilities in numerous Artificial Intelligence (AI) applications. Despite the increasing number of such AI HD applications, their susceptibility to adversarial attacks has not been explored, specifically in the text domain. To the best of our knowledge, this is the first research endeavour to evaluate the adversarial robustness of HD text classifiers and report on their vulnerability to such attacks. In this paper, we designed and developed n-grams based HD computing text classifiers for two primary applications of HD computing; language recognition and text classification, and then performed a set of character level and word level grey-box adversarial attacks, where an attacker’s goal is to mislead the target HD computing classifier to produce false prediction labels while keeping added perturbation noise as low as possible. Our results show that adversarial examples generated by the attacks can mislead the HD computing classifiers to produce incorrect prediction labels. However, HD computing classifiers show a higher degree of adversarial robustness in language recognition compared to text classification tasks. The robustness of HD computing classifiers against character-level attacks is significantly higher compared to word-level attacks and has the highest accuracy compared to deep learning-based classifiers. Finally, we evaluate the effectiveness of adversarial training as a possible defense strategy against adversarial attacks in HD computing text classifiers.

In this paper we introduce UNICON, a large-scale open dataset on UNIversity CONsumption of utilities, electricity, gas and water. This dataset is publicly released as part of La Trobe University’s commitment to Net Zero Carbon Emissions by 2029, for which we are building the La Trobe Energy AI/Analytics Platform (LEAP) that leverages Artificial Intelligence (AI) and Data Analytics to analyse, predict and optimize the consumption, generation and utilization of electricity, renewables, gas and water resources. UNICON contains consumption data for La Trobe’s five campuses in geographically distributed regions, across four years, 2018-2021 inclusive. This includes the COVID-19 global pandemic timeline of university shutdown and work from home measures that led to a significant decrease in the consumption of utilities. The consumption data consists of smart electricity meter readings at 15-minute granularity, gas meter readings at hourly intervals and water meter readings at 15-minute intervals. UNICON also contains weather data from the closest weather station to each campus, collected at two-speed latency of 1 minute and 10 minutes. The dataset is annotated with internal events of significance, such as energy conservation measures (ECMs) and other measurement and validation (M&V) activities conducted as part of LEAP optimization. To the best of our knowledge, this is the first large-scale, comprehensive, open dataset for the three main utilities, electricity, gas, and water consumption in a multi-campus university setting. A high granularity data dictionary and technical validation of the dataset for consumption trends, baseline modelling and forecasting are further contributions of this article that will enable interested research scientists, academics, industry practitioners, sustainability and energy consultants to experiment and evaluate their AI algorithms, models, forecasts, as well as inform the development of energy benchmarks, guidelines and much needed data-driven energy policies.

Recent advancements in Artificial intelligence (AI) have led to its widespread adoption in healthcare applications and services. The global pandemic has further hastened the integration of AI into telehealth services, such as service quality improvement and new models of care. This paper focuses on the service quality improvement of telehealth, specifically, cancer information and telephone support services, in terms of its human system interaction. We present an AI framework for the detection of patient emotions and emotion transitions based on call recordings of telehealth services. The call recordings are typically a conversation between the caller (patient or carer) and the healthcare practitioner (nurse or counsellor). All primary emotions are expressed across diverse topics during these calls. It is anticipated that the caller emotion state improves during the call due to the information and support received. The proposed AI framework is designed to detect emotions expressed at all stages of the call and based on these expressions, formulate the emotion transition during the call. We have evaluated the proposed AI framework on a large real-world dataset of 60,000 call recordings of cancer information and telephone support services provided by Cancer Council Victoria, Australia. The results confirm the effectiveness of this AI framework in detection of emotions and emotion transitions during the provision of telehealth services.

Time series forecasting is a critical requirement for the optimal operation of energy grids, systems, and platforms, where the forecasting challenge itself can span across energy consumption, renewables generation, and energy utilisation. Artificial Intelligence (AI) algorithms and models have been leveraged to predict these time series forecasts with increasing levels of accuracy. In contrast to local models that are developed separately for each time series, Global Models, which are trained across many sets of time series drawing on characteristics of ’relatedness’, have produced more accurate forecasts. In this paper, we propose a transformer architecture based global model as a generalist forecaster of energy time series data, where we frame a sequence forecasting model and represent numerical values of the corresponding time series as vector embeddings in this model. We evaluate this transformer architecture based global model on real-world time-series energy data generated by the La Trobe Energy AI platform (LEAP), a functional and operational microgrid deployed in the multicampus tertiary education setting of La Trobe University, Australia. The results of these experiments confirm that the proposed generalist forecasting approach outperforms specialist local models trained on individual time series. We also demonstrate the ability of this approach to forecast dissimilar time series from the same model.