İsMail Can Dikmen

This paper presents a physics guided learning framework for forecasting available traction power (ATP) and quantifying in motion charging (IMC) headroom in catenary fed trolleybus networks. The method combines a chronologically split, leak free feature policy, monotone gradient boosting with regime aware experts, an LSTM sequential baseline, and conformalized quantile regression for calibrated prediction intervals. Two case studies are reported. The first uses 58 345 high frequency samples from a dual power trolleybus in Malatya, Türkiye, where the monotone LightGBM reaches R² = 0.998 with mean absolute error around 1.3 kW and workload weighted error below 0.24% under a chronological split. The second uses the ZTBus dataset from Zurich, Switzerland, covering 1 409 missions, two vehicles, and 3.5 years of service. On the Zurich data, an LSTM baseline outperforms the tree based models (R² = 0.956, RMSE = 13.9 kW versus 17.0 kW for constrained LightGBM), while on the single run Malatya scenario the tree models dominate. A monotone constraint ablation shows that physics guided regularization slightly reduces in distribution accuracy but improves cross vehicle generalization. Conformal calibration delivers 90% prediction intervals with 93% empirical coverage on the Malatya dominant band and 79 to 81% on ZTBus under temporal distribution shift. Using the calibrated lowerATP quantile as a conservative headroom estimate, a 20 to 100 kW in motion charger recovers more than 99.8%of the oracle energy over the observed 31.0 km Malatya trip. These results show that the framework generalizes across cities, vehicles, and target variables, and can support in motion charging assessment in trolleybus networks.

This article presents the mathematical foundations of spiking neural networks (SNNs) in a unified formalism, with a deliberate emphasis on derivational provenance. The same neuron model is written one way in computational neuroscience textbooks, another way in machine learning papers, and a third way in the stochastic process literature. Even within a single line of work, papers absorb constants into other constants until two equations from two sources cannot be compared by inspection. We collect the core mathematics in one place, and we attach a status label to every major equation so that the reader sees at a glance whether a given step is a mathematical identity, a parameter limit, a formal approximation under stated conditions, or a useful but unproven heuristic. The labels are exact, reduction, approximation, and heuristic. The substantive content is the following. The reduction chain from Hodgkin-Huxley dynamics through the adaptive exponential integrate-and-fire model down to leaky integrate-and-fire (LIF) is given with status labels at every step, including the spike response model as an exact reformulation under linear subthreshold dynamics. Reset semantics are analyzed in three forms (hard, soft, no reset), with implications for both spike statistics and gradient flow. Network dynamics are written down in a coupled form, and the analytical theory of recurrent SNNs (liquid state machines, the echo state property, balanced excitatory-inhibitory networks) is reviewed with explicit conditions on time constants and weight matrices. The full point process formulation is developed: counting processes, conditional intensities, the time-rescaling theorem, the likelihood for general history-dependent point processes, and the canonical model classes (homogeneous Poisson, inhomogeneous Poisson, Hawkes, point-process generalized linear models). The bridge between state-space SNNs and intensity-based formulations is made explicit, including conditions under which a generalized linear model can be embedded in a finite-dimensional spiking state space. Information-theoretic aspects of spike coding are presented through Fisher information, with a quantitative comparison of rate and time-to-first-spike codes. Computational capacity is treated through three lines of results: the Maass third-generation argument and its noisy temporal-coding strengthening, the Stanojevic exact mapping from feedforward ReLU networks to time-to-first-spike SNNs, and the Date-Schuman Turing-completeness construction. The article closes with a status-labeled taxonomy of the hazard-based H-LIF family and its Liquid extension, drawn from a public, patent-scoped reference implementation with custom CUDA kernel and FPGA validation; the other LIF variant families (multi-spectral, wavelet, fractional, control-theoretic, information-theoretic, and domain-specific gating) are deferred to a companion v2.This article is the second installment in a series on spiking neural networks. The first installment, Spiking Neural Networks: A Tutorial on Models, Coding, and Training [1], introduces the practical side at a tutorial level; the present article develops the underlying mathematics in depth. The two share notation, and a reader who has followed the first installment can read this one essentially in any order; the cross-references between them are explicit. The intended audience is the graduate student or researcher who needs the mathematical underpinnings of SNNs in a single document, rather than reconstructed from a dozen textbooks and review papers.

In this study, the state of health (SoH) estimation of lithium-ion batteries (LIBs), which have become an indispensable part of modern life due to their widespread use in portable electronics and electric vehicles (EVs), has been conducted. Experimental data have been collected from three different battery chemistries—lithium titanate oxide (LTO), lithium manganese oxide (LMO), and lithium iron phosphate (LFP)—all in 18650 cylindrical format. These cells have been aged through charge–discharge cycling under constant current–constant voltage (CC–CV) conditions using a battery analyzer. Specifically, 2000, 400, and 1000 cycles have been completed for LTO, LMO, and LFP cells, respectively. The acquired datasets have been extracted, cleaned, and pre-processed, then subjected to differential voltage analysis (DVA), a data-driven technique widely utilized for SoH assessment. The DVA has proven effective in identifying detailed degradation patterns by analyzing subtle voltage changes during cycling. In this context, machine learning (ML) algorithms—namely Linear Regression (LR), Support Vector Machine (SVM), and Gaussian Process Regression (GPR)—have been integrated with DVA to improve the predictive accuracy of SoH estimation. This study considers SoH estimation as a regression task, and the combined use of DVA and ML represents a central novelty of the work. Based on regression analyses conducted in MATLAB, LR has provided the most accurate results for LTO and LMO, while the Matern 5/2 kernel-based GPR model has performed best for LFP. The corresponding RMSE values—1.5005 (LTO), 1.246 (LMO), and 1.667 (LFP)—reflect the predictive precision of each approach. These results demonstrate that SoH estimation has been successfully carried out with low error margins. The study confirms that combining DVA with appropriate regression models yields SoH predictions exceeding 98% prediction performance. This highlights the feasibility of data-driven, chemistry-specific diagnostics in battery health monitoring for EVs and ESSs. All results have been comparatively evaluated. Cite this article as: N. Yıldıran, İ. C. Dikmen and T. Karadağ, “Comprehensive modeling and prediction of state of health for 18650 lithium titanate oxide, lithium manganese oxide, and lithium iron phosphate batteries using machine learning algorithms,” Electrica, 26, 0028, 2026. doi: 10.5152/electrica.2026.25028.

Spiking neural networks (SNNs) are often called the third generation of neural models. They communicate with brief asynchronous pulses rather than continuous values, which suits event-driven sensors and low-power neuromorphic hardware. The mathematics behind them is split across neuroscience textbooks, machine learning papers, and stochastic process literature, and a researcher entering the area runs into a notation problem before anything else. The same neuron model is written one way in a textbook, another way in a machine learning paper, and a third way in the stochastic process literature. This tutorial collects what a graduate student or research engineer needs in order to start working with SNNs, in one place and with one notation. We start with the leaky integrate-and-fire (LIF) neuron, derived from a conductance-based picture of the membrane. Reset semantics, the spike response model, and the broader family that includes Hodgkin-Huxley and adaptive exponential models are discussed. Network equations are written out for feedforward and recurrent architectures. We cover the main neural coding schemes (rate, time-to-first-spike, rank-order, phase, burst, population) and discuss when each is appropriate. The neuromorphic datasets a beginner is likely to encounter, including N-MNIST, DVS-Gesture, CIFAR10-DVS, SHD, and SSC, are described together with the tensor formats event-based data takes. For learning, we derive pair-based spike-timing-dependent plasticity from exponential traces and develop the surrogate gradient framework, which has become the dominant tool for training deep SNNs by backpropagation through time. Reset semantics, the choice of surrogate function, and common pitfalls in BPTT are addressed in a way that maps onto code. Alternative training paradigms (e-prop, EventProp, SLAYER, three-factor rules, ANN-to-SNN conversion) are introduced briefly so the reader knows what else is available. A practical section walks through a first SNN training loop with framework-agnostic pseudocode and points to the main software libraries (snnTorch, SpikingJelly, Norse). Throughout the article, equations carry derivational status labels (exact, reduction, approximation, heuristic) so that the reader sees at a glance which steps are mathematical identities and which involve approximations. We do not cover hardware implementation, detailed point process theory, expressivity proofs, or open problems in SNN complexity; these belong in a more advanced treatment. The article is meant to be read linearly, and a suggested reading path closes it.

Spiking neural networks are increasingly considered for feedback control in energy constrained platforms, but converting established fuzzy controllers into event driven implementations while keeping evaluation transparent remains difficult. This paper presents two spiking backends for Takagi-Sugeno neuro-fuzzy controllers: a rate coded backend that encodes rule firing strengths as Bernoulli spike trains, and a sigma-delta backend that transmits only changes in firing strengths between consecutive control steps. A benchmarking workflow covers both open-loop inference fidelity and closed-loop control performance. In the open-loop track, spiking approximation error is measured against deterministic Sugeno inference. In the closed-loop track, spiking controllers are exercised on DC motor and cart-pole plants, where tracking accuracy is compared against PID and LQR baselines. A Takagi-Sugeno gain-scheduled architecture evaluated on the cart-pole achieves mean absolute error of [[EQUATION]]0.0113±0.0012~rad with [[EQUATION]]10.0±0.4 spikes per sample under rate coding, matching LQR performance ([[EQUATION]]0.0119±0.0013~rad). The sigma-delta backend reduces total spike cost by [[EQUATION]]4.7× at equivalent accuracy by exploiting temporal redundancy in closed-loop control signals, and achieves near-zero spike activity in steady-state. A theoretical analysis confirms that rate-coded approximation error follows [[EQUATION]] convergence while sigma-delta error is bounded by the quantization threshold. Computational cost projections for Loihi and SpiNNaker show that the sigma-delta backend extends the energetically favorable regime beyond the narrow sparse corner where rate coding is competitive. All experiments enforce tuning versus evaluation seed separation, and publish mode emits machine readable artifacts.

Large language models are increasingly consulted for scientific and clinical questions, yet ungrounded answers still appear too often to trust them on their own. We built a retrieval-augmented assistant that keeps generation tied to a curated, versioned corpus, and records every step from ingestion to answer. Documents are segmented with a practical, token-aware policy and encoded locally; vectors are stored with provenance so the system can cite or abstain. Queries are embedded, top-k passages are retrieved from a vector store, and a prompt asks the generator to respond only with supported statements or to decline. The components are intentionally swappable: the embedder runs on-premises for privacy, the store supports snapshots for repeatable experiments, and the generator (Gemma/Gemma2) is selected for efficient inference. Beyond the pipeline, we preregister an evaluation plan that measures retrieval quality, answer faithfulness and coverage, with ablations on chunk size, overlap, and k. All code, defaults, and scripts are released so others can reproduce the setup, compare their own choices, and extend the system to new domains. The goal is clear: reduce hallucination by grounding answers in literature, keep costs and latency predictable on a single-GPU server, and make empirical evaluation routine rather than optional. Experimental evaluation confirmed these design claims: the proposed modular RAG achieved Recall@k = 0.86, F1 = 0.79, and Attribution Accuracy = 0.91, significantly outperforming both Classic RAG and LLM-only baselines (p < 0.05). These results validate the framework’s reliability, grounding fidelity, and reproducibility for evidence-based clinical decision support.

Automotive edge systems face a growing gap between computational demand and what vehicle platforms can supply under tight power and thermal budgets, especially in autonomous vehicles. Neuromorphic computing is proposed as response, owing to its event driven operation. But earlier reviews on this subject tend to mix demonstrated uses with speculative applications and do not always relate efficiency claims to real driving conditions. This review addresses this gap in the literature; automotive system integration of neuromorphic hardware, spiking neural network training and deployment, event based sensing. Reviewed studies are separated into demonstrated implementations with measurable outcomes on stated platform and proposed opportunities that still lack automotive grade validation. Four observations are obtained from this review. First, efficiency gains from spike based processing become credible mainly when the workload is sparse and temporal by nature and when coding policy is selected with bounded time to decision in mind. Second, cross study comparison remains difficult because latency, energy, event rate condition, and stopping rule are usually reported in inconsistent ways across published studies. Third, deployment barriers are largely procedural, including toolchain maturity, integration of asynchronous accelerators with synchronous ECU timing, and the construction of safety arguments under ISO 26262 and SOTIF. Fourth, public industrial activity is still concentrated on bounded functions such as driver monitoring, keyword spotting, and radar pre processing rather than full neuromorphic autonomy stacks. Based on these findings, a deployment roadmap is proposed around always on modules with explicit timing contracts, automotive grade benchmark suites, and safety case patterns that constrain learning and enforce monitorable behavioral contracts.

Reliable state‑of‑charge estimation remains difficult in the presence of long‑term ageing, open‑circuit‑voltage hysteresis, and variable operating profiles. We investigate a hybrid estimator that marries a first‑principles equivalent‑circuit model with sequence learners, coupled through a residual‑aware fusion rule. Using a minute‑resolution ageing dataset from a lithium‑titanate (LTO) cell (236,282 samples across 2500 cycles), we identify phase‑aware OCV‑to‑SoC lookup tables for charge and discharge, deploy an extended Kalman filter with soft half‑cycle anchoring, and train LSTM, 1D‑CNN, and Transformer baselines with physics‑informed regularization promoting Coulomb consistency and concordance between OCV and terminal voltage. The phase‑aware EKF attains MAE 0.04535 and RMSE 0.05057 on cycle‑wise averages. A stitched LSTM yields MAE 0.03023 and RMSE 0.05192. Residual‑weighted fusion of EKF and LSTM produces MAE 0.03049 and RMSE 0.03991, which represents a 21% reduction relative to EKF while preserving the LSTM’s low bias. A coarse parameter sweep over the circuit model confirms expected early‑life behavior, namely lower resistance and a longer polarization time constant. At the window level, a physics‑informed LSTM achieves MAE 0.0184. The fusion is model‑agnostic, requires no retraining of the base estimators, and adds negligible computational overhead. We release phase‑aware OCV tables, a trained LSTM, and configuration files for seamless one‑command reproduction.

Bu çalışmada, Türkiye’nin Malatya kentinde hizmet veren troleybüs sisteminin elektriksel ve dinamik davranışları MATLAB/Simulink ortamında modellenmiş ve analiz edilmiştir. Modelde çekiş motorları, eviriciler, troley hattı, doğrultucu merkezleri ve gerilim düşümlerini temsil eden bileşenler ayrıntılı biçimde ele alınmıştır. Gerçek 36 km’lik hat verileri kullanılarak yapılan simülasyonlarda, hat boyunca oluşan gerilim ve akım değişimleri gözlemlenmiş, sistemin 750 V DC besleme altında nominal koşullarda çalıştığı doğrulanmıştır. Elde edilen sonuçlar, modelin gerçeğe yakın davranış sergilediğini ve troleybüs sistemlerinin enerji verimliliği analizlerinde güvenle kullanılabileceğini göstermektedir. Bu kapsamda çalışma, kentsel elektrikli ulaşım sistemlerinin tasarım, planlama ve optimizasyon süreçlerinde referans oluşturabilecek niteliktedir.