An Agentic AI Pipeline for Appliance-Level Energy Anomaly Detection and LLM-Driven Recommendations

# An Agentic AI Pipeline for Appliance-Level Energy Anomaly Detection and LLM-Driven Recommendations
Source: [https://arxiv.org/html/2606.28467](https://arxiv.org/html/2606.28467)
###### Abstract

Appliance\-level energy monitoring in office buildings produces noisy alerts that non\-expert facility managers struggle to use\. This paper proposes an end\-to\-end agentic pipeline that combines deep time\-series forecasting, variational anomaly detection, and LLM\-based reasoning to generate prioritized, actionable maintenance recommendations\. The system tracks seven office appliances using a hybrid Singular Spectrum Analysis \(SSA\) and Long Short\-Term Memory \(LSTM\) forecasting model, and applies a per\-appliance LSTM Variational Autoencoder \(VAE\) with attention to flag abnormal daily consumption episodes\. A three\-stage LangChain pipeline begins with a Context Agent that always retrieves three core RAG sources model reliability, hourly baseline, and expert knowledge and conditionally adds up to three more \(forecast context, anomaly history, global baseline\) based on event characteristics, capped at eight reasoning steps\. A Diagnosis Agent converts the evidence into a structured JSON diagnosis, and a Report Agent renders a human\-readable narrative\. A reflective memory layer incorporates operator feedback\. The dashboard shows real\-time 30\-minute forecasts, intraday consumption, the previous day anomaly report, and a feedback form\. We evaluate the forecasting model, anomaly detector with appliance\-specific thresholds, and LLM reasoning on a 16\-scenario benchmark including sustained and transient spikes, unexpected shutdowns, and systemic events, comparing five LLM backends under static vs\. dynamic retrieval\. Dynamic retrieval matches full static retrieval across all backends while cutting average context from six to three–six sources per event\. The best backend scores 90\.4/100 with a 100% pass rate at a 70\-point threshold, and a fully local 7B\-parameter model passes all 16 scenarios\.

††publicationid:pubid:© 2025 IEEE© 2025 IEEE\. This is the author’s version of the work accepted for publication in the IEEE International Conference on Communication, Computing, Networking, and Control in Cyber\-Physical Systems \(CCNCPS 2026\)\. The final version will be available via IEEE Xplore\.

## IIntroduction

Buildings account for a substantial share of global electricity demand, and office environments in particular suffer from waste caused by faulty appliances, forgotten devices, and degraded equipment that continue to draw power without being noticed\[[1](https://arxiv.org/html/2606.28467#bib.bib1)\]\. Sub\-metering technologies now make it routine to record energy consumption at the appliance level, yet most facility managers still rely on aggregated dashboards and manual inspection to react to anomalies\[[2](https://arxiv.org/html/2606.28467#bib.bib2)\]\. The problem is no longer one of measurement, but one of*interpretation*: a stream of half\-hourly readings from a fridge, a coffee machine, and a printer must be turned into a small number of clear, prioritized actions\. Deep learning \(DL\) approaches for time\-series forecasting and anomaly detection now enable reliable identification of unusual appliance behavior without labeled fault data\[[3](https://arxiv.org/html/2606.28467#bib.bib3),[4](https://arxiv.org/html/2606.28467#bib.bib4)\]\. Hybrid decomposition models such as SSA\-LSTM\[[8](https://arxiv.org/html/2606.28467#bib.bib8)\]improve forecasting accuracy, while LSTM\-VAEs with percentile\-based thresholds\[[9](https://arxiv.org/html/2606.28467#bib.bib9)\]convert reconstruction errors into appliance\-level alerts\. However, global thresholds fail for heterogeneous appliance portfolios\[[9](https://arxiv.org/html/2606.28467#bib.bib9)\], and most systems stop at alert generation, placing the interpretive burden on facility managers who lack the technical expertise to prioritize raw anomaly scores\[[2](https://arxiv.org/html/2606.28467#bib.bib2)\]\. Recent work addresses interpretability by encoding energy time series as wavelet or recurrence\-plot images and feeding them to vision\-LLMs\[[17](https://arxiv.org/html/2606.28467#bib.bib17),[18](https://arxiv.org/html/2606.28467#bib.bib18)\], demonstrating that visual representations improve diagnostic fidelity; however, these end\-to\-end architectures sacrifice the modularity needed for safety\-critical deployments\. At the feature extraction level, attention\-based transformer representations have shown strong capacity for capturing temporal dependencies in time\-series anomaly detection\[[13](https://arxiv.org/html/2606.28467#bib.bib13)\], motivating their use as an interpretable intermediate layer before alert generation\. On the reasoning side, ReAct\-style agents\[[5](https://arxiv.org/html/2606.28467#bib.bib5)\]combine LLM reasoning with external tool use, while retrieval\-augmented generation \(RAG\)\[[6](https://arxiv.org/html/2606.28467#bib.bib6)\]grounds outputs in external knowledge bases implemented with FAISS\[[14](https://arxiv.org/html/2606.28467#bib.bib14)\]and lightweight encoders such as MiniLM\[[15](https://arxiv.org/html/2606.28467#bib.bib15)\]\. LangChain\[[10](https://arxiv.org/html/2606.28467#bib.bib10)\]supports structured tool orchestration, while recent surveys highlight increasingly sophisticated behaviors of LLM agents\[[11](https://arxiv.org/html/2606.28467#bib.bib11)\]\. These agents have been applied to building energy management tasks such as device control, scheduling, and energy analysis\[[12](https://arxiv.org/html/2606.28467#bib.bib12)\]\. Adaptive retrieval strategies\[[20](https://arxiv.org/html/2606.28467#bib.bib20)\]further show that selecting context conditionally based on query characteristics matches full\-retrieval accuracy while reducing overhead\. Nevertheless, LLMs risk hallucination and overconfident urgency assignments when deployed without deterministic safety nets\[[7](https://arxiv.org/html/2606.28467#bib.bib7)\], and agentic pipelines that fix the set of retrieved sources at every call incur unnecessary context overhead for simple events\. These limitations motivate the present work\. We propose Smart Energy Agent, a hybrid agentic pipeline that combines \(i\) an SSA\-LSTM forecasting backbone shared across appliances, \(ii\) per\-appliance LSTM\-VAE with multi\-head attention anomaly detectors with calibrated thresholds, and \(iii\) a three\-stage LLM agent that retrieves three core RAG sources unconditionally and conditionally selects up to three additional sources based on event characteristics, combined with a feedback memory, before producing structured diagnoses and natural\-language recommendations\. Hard rules and reliability caps surround the LLM to prevent overconfident or unsafe outputs\. The contributions of this work are:

- •A unified architecture coupling deep forecasting, deep anomaly detection, and an LLM agent through retrieval\-augmented reasoning, with explicit safety nets on urgency and confidence\.
- •A two\-tier retrieval strategy in which the Context Agent always retrieves three core RAG sources \(model reliability, hourly baseline, and expert knowledge\) and conditionally selects up to three additional sources \(forecast context, anomaly history, and global baseline\) based on event characteristics, reducing average context size from 6 to 3–6 sources per event\.
- •A per\-appliance threshold calibration strategy that adapts detector strictness to the noise profile of each device\.
- •A 16\-scenario evaluation benchmark with a 100\-point rubric covering urgency classification, likely cause, confidence calibration, special\-case flags, and actionability, used to compare dynamic versus static retrieval across five LLM backends, confirming equivalent diagnostic accuracy with reduced context overhead\.
- •A comparison of five LLM backends including a fully local 7B\-parameter model, confirming pipeline viability without external API dependencies\.

## IIMethodology

Fig\.[1](https://arxiv.org/html/2606.28467#S2.F1)summarizes the end\-to\-end architecture of the proposed system, from raw sensor readings through forecasting, anomaly detection, and agentic reasoning to the operator dashboard\. The architecture operates at two 11timescales\. Every 30 minutes, sensor readings feed the forecasting module, which appends predictions and actuals toforecast\_log\.csvand updates the live monitoring view\. Every 24 hours, the anomaly detector scores all appliances, overwritesanomaly\_events\.jsonwith enriched event objects, and appends new flags toanomaly\_history\.csv; the Context Agent, Diagnosis Agent, and Report Agent then consume these outputs to produce the daily report\.

### II\-ASensors and Data Layer

Seven appliances are monitored — a coffee machine, a fridge, a microwave, a kettle, a printer, a water dispenser, and a Tasmota\-based smart plug — each sampled every 30 minutes as a \(timestamp, appliance, kWh\) tuple\. Four persistent stores are maintained:raw\_readings\.csv,forecast\_log\.csv\(seven rows are appended for each appliance per forecasting run\),anomaly\_history\.csv\(a rolling 30\-day flag record updated nightly\),anomaly\_events\.json\(enriched event objects overwritten each nightly run\), andfeedback\_log\.csv\(operator corrections\)\.

### II\-BComponent Overview

- •Forecasting Model:SSA\-LSTM hybrid producing the next 30\-minute prediction per appliance; outputs a*surprise ratio*used downstream\.
- •Anomaly Detector:Seven independent LSTM\-VAE models with multi\-head attention and percentile\-based thresholds\.
- •Coordinator:Groups events by time and appliance, filters borderline cases, and caps LLM calls per run\.
- •Context Agent:Tool\-calling executor with up to eight reasoning iterations; applies two\-tier RAG retrieval\.
- •Diagnosis Agent:Applies hybrid rule \+ LLM logic to produce a structured JSON diagnosis\.
- •Report Agent:Converts JSON diagnoses into a prioritized natural\-language report\.
- •Feedback Memory:Reflective store of past operator corrections injected into future diagnosis prompts\.
- •Dashboard:Next\.js/React application with live monitoring, daily report, and feedback form views\.

### II\-CForecasting with SSA\-LSTM

The forecasting backbone is a single global model with separate LSTM branches for each SSA component \(trend, seasonality, residual\) and an additional branch for exogenous features\. A learned per\-appliance embedding allows specialization without per\-device training\. Inputs include lagged consumption values, rolling statistics, and cyclic time encodings\. The model is trained once on a chronological 70/15/15 split and deployed in inference\-only mode every 30 minutes\.

![Refer to caption](https://arxiv.org/html/2606.28467v1/fig_sys_1.png)Figure 1:End\-to\-end architecture of the Smart Energy Agent\.
### II\-DAnomaly Detection with LSTM\-VAE

For each appliance we train a separate LSTM\-VAE: a bidirectional LSTM encoderfϕf\_\{\\phi\}with multi\-head attention produces an approximate posteriorqϕ​\(z∣x\)=𝒩​\(μϕ​\(x\),σϕ2​\(x\)\)q\_\{\\phi\}\(z\\mid x\)=\\mathcal\{N\}\(\\mu\_\{\\phi\}\(x\),\\sigma\_\{\\phi\}^\{2\}\(x\)\), and an LSTM decodergθg\_\{\\theta\}reconstructs the input windowx∈ℝT×Fx\\in\\mathbb\{R\}^\{T\\times F\}\. Training minimizes a feature\-weighted ELBO,

ℒ​\(ϕ,θ;x\)=𝔼qϕ​\(z\|x\)​\[1T​F​∑t,fwf​\(xt,f−gθ​\(z\)t,f\)2\]\+β​DKL​\(qϕ​\(z\|x\)∥𝒩​\(0,I\)\),\\begin\{split\}\\mathcal\{L\}\(\\phi,\\theta;x\)&=\\mathbb\{E\}\_\{q\_\{\\phi\}\(z\|x\)\}\\\!\\left\[\\tfrac\{1\}\{TF\}\\textstyle\\sum\_\{t,f\}w\_\{f\}\\bigl\(x\_\{t,f\}\-g\_\{\\theta\}\(z\)\_\{t,f\}\\bigr\)^\{2\}\\right\]\\\\ &\\quad\+\\beta\\,D\_\{\\mathrm\{KL\}\}\\\!\\left\(q\_\{\\phi\}\(z\|x\)\\,\\\|\\,\\mathcal\{N\}\(0,I\)\\right\),\\end\{split\}\(1\)wherewfw\_\{f\}up\-weights energy features \(e\.g\.wkWh=10w\_\{\\mathrm\{kWh\}\}\{=\}10\) andβ\\betais annealed from0to0\.050\.05to prevent posterior collapse\. At inference, a sliding window scores each timestep by averaging the weighted reconstruction error over all overlapping windows that contain it, usingμϕ\\mu\_\{\\phi\}as a deterministic latent code,

s​\(t\)=1\|𝒲​\(t\)\|​∑w∈𝒲​\(t\)1F​∑fwf​\(xt,f−gθ​\(μϕ​\(x\(w\)\)\)t,f\)2\.s\(t\)=\\frac\{1\}\{\|\\mathcal\{W\}\(t\)\|\}\\sum\_\{w\\in\\mathcal\{W\}\(t\)\}\\tfrac\{1\}\{F\}\\textstyle\\sum\_\{f\}w\_\{f\}\\bigl\(x\_\{t,f\}\-g\_\{\\theta\}\(\\mu\_\{\\phi\}\(x^\{\(w\)\}\)\)\_\{t,f\}\\bigr\)^\{2\}\.\(2\)The detection thresholdτ\\tauis thepp\-th percentile of validation scores \(chronological middle 15%, used solely for calibration\),

τ=Qp​\(\{s​\(t\):t∈𝒱\}\),\\tau=Q\_\{p\}\\\!\\left\(\\\{s\(t\):t\\in\\mathcal\{V\}\\\}\\right\),\(3\)withpptuned per appliance: Microwave 97\.0%, Printer 98\.5%, Fridge/Kettle/Coffee Machine 99\.0%, Water Dispenser 99\.5%, and Tasmota 99\.9%\. Noisier devices receive lower thresholds to preserve sensitivity; more predictable ones receive stricter thresholds since a high reconstruction error is genuinely surprising\. Consecutive flagged timesteps are merged into events and enriched with energy statistics, severity, and pattern type\.

### II\-EAgentic Pipeline

The pipeline comprises three sequential stages: a Context Agent for evidence retrieval, a Diagnosis Agent for interpretation, and a Report Agent for narrative generation\. The Context Agent is a LangChain tool\-calling executor with six registered tools:

1. 1\.get\_expert\_knowledge\(\)– semantic search over a FAISS\-backed appliance knowledge base usingall\-MiniLM\-L6\-v2embeddings\.
2. 2\.get\_model\_reliability\(\)– the false\-alarm profile of the detector for that appliance\.
3. 3\.get\_hourly\_baseline\(\)– expected consumption for the specific hour and weekday\.
4. 4\.get\_global\_baseline\(\)– training\-set energy statistics for the appliance\.
5. 5\.get\_anomaly\_history\(\)– recent anomalies for that appliance\.
6. 6\.get\_forecast\_context\(\)– whether the forecaster was also surprised at the time of the event\.

The first three are retrieved unconditionally — the agent must always know detector reliability, the time\-of\-day baseline, and domain knowledge before reasoning\. The remaining three are conditionally selected based on signal ambiguity, event severity, or baseline sparsity, keeping the diagnosis prompt focused and free of irrelevant context\. The Diagnosis Agent routes each event through hybrid rule \+ LLM logic across three paths: Path 1 returns a deterministic “inspect now” diagnosis for devices completely inactive for several days; Path 2 closes the event as normal variation if statistics fall within the historical range; Path 3 invokes the LLM with the full retrieved context, producing a JSON object with eight fields:diagnosis,likely\_cause,confidence,confidence\_reason,recommended\_action,urgency,urgency\_reason, andexplanation\. Four post\-generation safety nets are applied: an urgency fallback, a confidence fallback, a reliability cap that lowers confidence for known noisy detectors, and a night\-time hard rule that forcesimmediateurgency for fridge or water dispenser anomalies between 10 PM and 6 AM\. If three or more appliances are simultaneously anomalous, a systemic event rule redirects the diagnosis to a building\-level cause\. The Report Agent converts the structured JSON into a prioritized Markdown report grouped by urgency\.

### II\-FFeedback Loop and Dashboard

Operator corrections are appended tofeedback\_log\.csvvia the dashboard feedback form\. On each nightly run, corrections matching the current event by appliance, pattern type, and time\-of\-day window are injected into the diagnosis prompt, providing a reflective memory that progressively reduces repeated false alarms\. The dashboard provides three views: live monitoring \(full intraday consumption trace per appliance with the next 30\-minute forecast, refreshed each sensor cycle\), daily report \(yesterday’s anomalies rendered as a structured narrative with severity, urgency, likely cause, recommended action, and confidence\), and a feedback form for operator corrections\.

## IIIExperimental Evaluation

### III\-ASetup

All experiments are conducted on a chronological train/validation/test split of the appliance\-level dataset\. The validation set \(the middle 15%\) is used solely for threshold calibration; metrics are reported on the held\-out test set\. The forecasting model uses 30\-minute granularity\. The LLM stage is evaluated separately on a benchmark of 16 hand\-crafted scenarios \(S01–S16\) covering sustained spikes, transient spikes, unexpected shutdowns, extended inactivity, high\-variability episodes, and systemic events across all seven appliances\.

### III\-BForecasting Performance

Table[I](https://arxiv.org/html/2606.28467#S3.T1)reports the global metrics across all appliances\. The model achieves anR2R^\{2\}of 0\.9976 and a Weighted Absolute Percentage Error \(WAPE\) of 1\.32%, indicating that variance is well captured at the aggregate level\. Table[II](https://arxiv.org/html/2606.28467#S3.T2)breaks results down per appliance\.

TABLE I:Global Forecasting Metrics on the Test SetTABLE II:Per\-Appliance Forecasting Metrics on the Test SetThe fridge and the printer, both with stereotypical patterns, are forecast almost perfectly\. The microwave shows the highest relative error \(WAPE = 21\.2%\), consistent with its bursty on/off behavior\. The water dispenser exhibits the largest absolute RMSE \(31\.84 kWh\), driven by infrequent but intense heating spikes\. These patterns motivate the appliance\-specific threshold calibration of the anomaly detector\.

### III\-CAnomaly Detection Reliability

Table[III](https://arxiv.org/html/2606.28467#S3.T3)summarizes the reliability profile of each LSTM\-VAE model\. Reliability is a qualitative trust assessment derived from threshold strictness, validation behavior, and observed false\-alarm rates over 30 days of operation\. Devices marked LOWER such as the microwave are flagged frequently and require corroborating evidence; devices marked HIGH \(fridge, printer\) are trusted by default when they raise an alert\.

TABLE III:Per\-Appliance Detector Reliability and Threshold
### III\-DLLM Backend Comparison

The LLM reasoning stage is evaluated using a 100\-point rubric per scenario, broken down as follows:

- •Urgency classification \(35 pts\): full / partial / zero match against the expected level\.
- •Likely cause keywords \(25 pts\): proportional, synonym\-aware coverage of the expected cause\.
- •Confidence level \(20 pts\): match between stated confidence and underlying model reliability\.
- •Special flags \(10 pts\): correct flagging of night\-time, concurrent, or reliability\-related caveats\.
- •Actionability \(10 pts\): presence of a concrete action verb and target object\.

A scenario is considered*passed*when its score reaches 70/100 or higher\. We tested five backends from open\-source \(Ollama\) and commercial APIs \(Anthropic, OpenRouter\)\. Each was evaluated under both static retrieval \(all six RAG sources retrieved unconditionally\) and dynamic retrieval \(three core sources always retrieved, up to three additional selected conditionally\)\. Table[IV](https://arxiv.org/html/2606.28467#S3.T4)shows dynamic retrieval results, which matched static scores exactly across all backends and scenarios\.

TABLE IV:Scenario\-Based Evaluation of Five LLM Backends \(16 Scenarios, Dynamic Retrieval\)Three observations stand out\. First, hosted models cluster within a 1\.9\-point range \(88\.5–90\.4\), suggesting frontier model choice has limited impact on diagnosis quality given sufficient retrieval context\. Second, all hosted models pass at least 15/16 scenarios, with only one borderline case separating Gemini\. Third, the local 7B Qwen2\.5 passes all 16 scenarios at 85\.4—roughly five points below the best hosted model but well above threshold, confirming API\-free feasibility\. Dynamic retrieval matched full static retrieval accuracy while reducing average context from 6 to 3–6 sources per event\. Although retrieval cost differences are negligible in local RAG setups, dynamic selection becomes cost\-justified in production environments with API\-based or rate\-limited sources, where skipping 2–3 retrievals per event compounds significantly across thousands of daily monitoring cycles\.

### III\-EComponent Ablation

The hybrid routing logic plays a measurable role in robustness\. Removing the night\-time hard rule, the reliability cap, or the rule\-based urgency fallback caused intermittent failures during exploratory development, particularly on rare or under\-represented event types\. The deterministic Path 1 short\-circuit \(devices completely off for several days\) further reduces unnecessary LLM calls and avoids spurious confidence values on what is, in fact, a certainty\.

### III\-FComparison with Related Work on LLM Reasoning Evaluation

Table[V](https://arxiv.org/html/2606.28467#S3.T5)maps the three most relevant cited systems onto a common set of dimensions\. Two observations follow\. The two closest architectural relatives\[[17](https://arxiv.org/html/2606.28467#bib.bib17),[18](https://arxiv.org/html/2606.28467#bib.bib18)\]evaluate LLM output via validation loss and AHP scoring respectively — neither decomposes diagnostic quality along the operationally critical dimensions of urgency correctness, confidence calibration, or actionability that our rubric captures\. Their results confirm that visual representations improve detection fidelity, but neither provides a pass/fail diagnostic benchmark comparable to ours\. Finally, RAGAS\[[16](https://arxiv.org/html/2606.28467#bib.bib16)\]and ARES\[[19](https://arxiv.org/html/2606.28467#bib.bib19)\]are evaluation*frameworks*rather than systems, and both confirm that generic automated metrics are insufficient for domain\-specific diagnostic tasks — directly motivating the hand\-crafted 100\-point rubric used here\.

TABLE V:Comparison of LLM Reasoning Evaluation Across Related WorkNotes: “Validation loss” \(lower↓\\downarrowis better\) and “AHP score” \(higher↑\\uparrowis better\) are not directly comparable to the rubric score used in this work; they are included to characterise each system’s evaluation methodology\.

## IVDiscussion

The most effective design choice was separating numerical evidence gathering from language generation: forecasting and anomaly detection produce structured artifacts; the Context Agent assembles a focused briefing; the Diagnosis Agent emits strict JSON; the Report Agent renders the narrative\. This localizes errors — a misformatted JSON does not corrupt the report, and a poorly worded report does not affect the structured event log\.

Per\-appliance thresholds yielded a clear payoff over a single global threshold, which either flooded operators with microwave alerts or suppressed genuine fridge anomalies\. Dynamic context selection confirmed that diagnostic performance was identical whether three or six sources were retrieved, validating focused retrieval — a benefit expected to compound in production environments with rate\-limited external APIs\. The reflective feedback memory is fully integrated, injecting operator corrections into future prompts when a matching event is found by appliance, pattern type, and time of day; quantifying its impact on accuracy requires sustained expert\-validated corrections and is left for future work\. The 16\-scenario benchmark covers the operationally relevant event types observed over 30 days; a larger adversarial benchmark would further strengthen evaluation\. The full stack runs on free\-tier infrastructure: FAISS withall\-MiniLM\-L6\-v2on CPU, with an optional local Qwen2\.5 backend for on\-premises data residency\.

## VConclusion

We presented Smart Energy Agent, a hybrid pipeline combining deep forecasting, anomaly detection, retrieval\-augmented LLM reasoning, and operator feedback for appliance\-level energy management\. Separating numerical evidence gathering from language generation and surrounding the LLM with deterministic safety nets yields reliable, prioritized recommendations even when detectors are noisy\. Forecasting reachesR2=0\.9976R^\{2\}=0\.9976on the held\-out test set; the agentic stage scores 90\.4/100 on a 16\-scenario benchmark with the best hosted backend, while a fully local 7B model passes all 16 scenarios — confirming viability without external API dependencies\. Future work will focus on three directions: extending the benchmark with adversarial scenarios to stress\-test the safety nets, quantifying the impact of the reflective feedback memory on repeated false\-alarm reduction over time, and validating dynamic retrieval cost savings under production\-scale deployments with external knowledge sources\.

## Acknowledgment

The authors thank the team at Laboratoire LITAN, ESTIN, for ongoing discussions on agentic AI systems for energy management\. The authors also thank the College of Computing and Informatics \(CCI\) at the University of Sharjah and the College of Engineering and Information Technology \(CEIT\) at the University of Dubai for their support and collaboration\.

## References

- \[1\]International Energy Agency, “Energy Efficiency 2023,” IEA, Paris, 2023\.
- \[2\]Y\. Himeur, K\. Ghanem, A\. Alsalemi, F\. Bensaali, and A\. Amira, “Artificial intelligence based anomaly detection of energy consumption in buildings: A review, current trends and new perspectives,” Applied Energy, vol\. 287, 116601, 2021\.
- \[3\]Z\. Zamanzadeh Darban, G\. I\. Webb, S\. Pan, C\. C\. Aggarwal, and M\. Salehi, “Deep learning for time series anomaly detection: A survey,”ACM Computing Surveys, vol\. 57, no\. 1, Article 15, Oct\. 2024\.
- \[4\]T\. Nguyen and V\. Tran, “Deep learning for multivariate time series anomaly detection: An evaluation of reconstruction\-based methods,”Artificial Intelligence Review, Springer Nature, 2025\.
- \[5\]S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao, “ReAct: Synergizing reasoning and acting in language models,” inProc\. ICLR, 2023\.
- \[6\]P\. Lewis et al\., “Retrieval\-augmented generation for knowledge\-intensive NLP tasks,” Proc\. NeurIPS, 2020\.
- \[7\]Z\. Ji et al\., “Survey of hallucination in natural language generation,” ACM Computing Surveys, vol\. 55, no\. 12, pp\. 1–38, 2023\.
- \[8\]N\. Moishin, R\. Deo, R\. Prasad, N\. Mohammed, and S\. Salcedo\-Sanz, “Short\-term net load forecasting with singular spectrum analysis and LSTM neural networks,”Energies, vol\. 14, no\. 14, p\. 4107, 2021\.
- \[9\]R\. Schmidl, P\. Wenig, and T\. Papenbrock, “Anomaly detection in time series: A comprehensive evaluation,”Proceedings of the VLDB Endowment, vol\. 15, no\. 9, pp\. 1779–1797, 2022\.
- \[10\]H\. Chase, “LangChain: Building applications with LLMs through composability,”https://github\.com/langchain\-ai/langchain, 2023\.
- \[11\]L\. Wang, C\. Ma, X\. Feng, Z\. Zhang, H\. Yang, J\. Zhang, Z\. Chen, J\. Tang, X\. Chen, Y\. Lin, et al\., “A survey on large language model based autonomous agents,” Frontiers of Computer Science, vol\. 18, no\. 6, 186345, 2024\.
- \[12\]Y\. Himeur, A\. N\. Sayed, A\. Alsalemi, F\. Bensaali, and A\. Amira, “Edge AI for internet of energy: Challenges and perspectives,”Internet of Things, vol\. 24, p\. 101035, 2023\.
- \[13\]J\. Kim, H\. Kang, and P\. Kang, “Time\-series anomaly detection with stacked transformer representations and 1D convolutional network,”Engineering Applications of Artificial Intelligence, vol\. 120, p\. 105–114, 2023\.
- \[14\]J\. Johnson, M\. Douze, and H\. Jégou, “Billion\-scale similarity search with GPUs,” IEEE Transactions on Big Data, vol\. 7, no\. 3, pp\. 535–547, 2021\.
- \[15\]W\. Wang et al\., “MiniLM: Deep self\-attention distillation for task\-agnostic compression of pre\-trained transformers,” Proc\. NeurIPS, 2020\.
- \[16\]S\. Es, J\. James, L\. Espinosa\-Anke, and S\. Schockaert, “RAGAS: Automated evaluation of retrieval augmented generation,” Proc\. EACL, 2024\.
- \[17\]A\. Bechar, A\. Oulefki, A\. Amira, F\. Kurugollu, and Y\. Himeur, “Extracting actionable insights from building energy data using vision LLMs on wavelet and 3D recurrence representations,” in Proc\. IEEE Int\. Conf\. Data Mining \(ICDM\), 2025, arXiv:2509\.21934\.
- \[18\]A\. Bechar, A\. Amira, A\. Oulefki, and Y\. Himeur, “Decoding energy consumption patterns through integration of visual encoding techniques and vision large language models,” Expert Systems with Applications, 2026\.
- \[19\]J\. Saad\-Falcon, O\. Khattab, C\. Potts, and M\. Zaharia, “ARES: An automated evaluation framework for retrieval\-augmented generation systems,” inProc\. NAACL, 2024\.
- \[20\]A\. Asai, Z\. Wu, Y\. Wang, A\. Sinha, and H\. Hajishirzi, “Self\-RAG: Learning to retrieve, generate, and critique through self\-reflection,” inProc\. ICLR, 2024\.