Koop, Karsjen (2026) Scalable Peer-to-Peer Energy Matching: A Dual-Tower Approach utilizing CoST Encoders, Projection Heads, and Soft-SigLIP Loss. Bachelor thesis, Data Science and Society (DSS).
|
PDF
BA4380304KKoop-1-1.pdf Download (2MB) | Preview |
Abstract
A potential solution to decentralized renewable energy is peer-to-peer (P2P) matching based on high-frequency 15-minute load profiles. However, optimal one-to-one pairings scale at an unsustainable quadratic complexity ( ). This research proposes a Dual-Tower neural network that frames energy matching as a cross-modal representation learning task. By compressing asymmetrical supply and demand time-series into a shared latent space, the framework enables sub-linear Approximate Nearest Neighbor (ANN) retrieval. Evaluating self-supervised CoST encoders against end-to-end 1D-CNNs, and Soft-BCE versus Soft-SigLIP loss, testing revealed that the network is highly susceptible to latent space collapse. This is mitigated by strictly regularizing the physical ground-truth using an asymmetric harmonic mean, identifying a universal stability window between and . Isolated within this stable regime, optimal configurations captured over 96% of the theoretical maximum physical energy synergy. A nuanced architectural trade-off emerged: the 1D-CNN achieved the highest absolute peak performance (54.16% improvement over a random baseline), while the disentangled CoST backbone demonstrated superior reliability (52.53% median peak) and rapid convergence. Additionally, Soft-SigLIP’s dynamic geometric fitting provided a significantly more robust operational ceiling compared to rigid Soft-BCE bounds. Ultimately, this scalable methodology proves high-frequency energy matching is computationally viable, laying the technical foundation for democratized decentralized grids.
| Item Type: | Thesis (Bachelor) |
|---|---|
| Name supervisor: | Nayak, S. |
| Date Deposited: | 30 Jul 2026 07:39 |
| Last Modified: | 30 Jul 2026 07:39 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/898 |
Actions (login required)
 |
View Item |