Research map / Time series / Long-term forecasting
Efficient and Effective Time-Series Forecasting with Spiking Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a framework that adapts CNN, RNN/GRU, and Transformer (iTransformer) backbones into spiking neural networks with delta and convolutional spike encoders for multivariate time-series forecasting, reporting comparable accuracy to ANNs with much lower estimated energy.
Problem
Applying spiking neural networks to time-series forecasting is hard due to temporal alignment between continuous signals and discrete spikes, the complexity of encoding floating-point values into spike trains, and lack of guidance on SNN model selection.
Approach
Align SNN time steps with time-series steps by subdividing each series step into Ts spiking sub-steps, use learnable convolutional or delta spike encoders instead of repetition, and convert TCN, RNN/GRU and iTransformer into hardware-friendly spiking counterparts (Spike-TCN, Spike-RNN/GRU, iSpikformer) trained with surrogate gradients.
Key results
- iSpikformer attains average RSE 0.476 across benchmarks versus 0.480 for iTransformer, with R^2 0.775 vs 0.776, i.e. comparable accuracy.
- Spike-RNN achieves average R^2 0.768 (significant at p<0.05 over RNN-style baseline per table) and average RSE 0.503.
- Spike-GRU, Spike-TCN and iSpikformer reduce estimated inference energy by 75.05%, 63.60% and 66.30% respectively on Electricity; average reduction about 70.33%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the ANN GRU with Spike-GRU | GRU · Four benchmarks, Avg. | R^2 | Helped |
| Replace the ANN GRU with Spike-GRU | GRU · Four benchmarks, Avg. | RSE | Helped |
| Replace the ANN iTransformer with iSpikformer | iTransformer · Four benchmarks, Avg. | RSE | Helped |
| Replace the ANN iTransformer with iSpikformer | iTransformer · Four benchmarks, Avg. | R^2: 0.001 | No clear effect |
| Use a convolutional spike encoder instead of a delta spike encoder (ablation) | Spike-TCN, Spike-RNN, and iSpikformer · Metr-la and Electricity | R^2: +0.09 on average | Helped |
| Use a repetition encoder instead of the convolutional spike encoder (ablation) | Spike-TCN, Spike-RNN, and iSpikformer · Metr-la and Electricity | R^2 | Hurt |
| Increase SNN time steps to 16 (ablation) | SNN models · Metr-la and Solar, L=24 | R^2 | Hurt |
| Increase decay rate beta (ablation) | SNN models · Metr-la and Solar, L=24 | R^2 | Hurt |
| Replace TCN with Spike-TCN | TCN · Four benchmarks, Avg. | R^2 | Hurt |
| Replace TCN with Spike-TCN | TCN · Four benchmarks, Avg. | RSE | Hurt |
Surprising findings
- Repetition encoders often fail to converge (R^2 near 0.02) for Spike-TCN on Metr-la, showing naive encoding breaks SNN training.
- Spike-TCN underperforms its ANN TCN counterpart, attributed to not tracking membrane state across time steps.
Limitations
- Energy figures are theoretical estimates from SOP/FLOP counts, not measured on neuromorphic hardware; GPU runs do not reflect SNN speedups.
- Regression output uses a non-spiking floating-point linear decoder, so the full network is not purely spiking.
- Spike-TCN requires resetting membrane potential each step, limiting temporal modeling and parallelism; spatial modeling is not addressed by spiking graph methods.
Experimental setup
Datasets: Metr-la Pems-bay Solar-energy Electricity synthetic periodic signals
Benchmarks: Metr-la Pems-bay Solar-energy Electricity
Metrics: R^2 RSE energy consumption (mJ) critical difference
Baselines: ARIMA GP TCN GRU Autoformer iTransformer
Builds on: iTransformer Spikformer TCN SEW ResNet snnTorch
Closest papers by meaning
- TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting (ICLR 2025 · similarity 0.90)
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.89)
- Learning Delays in Spiking Neural Networks using Dilated Convolutions with Learnable Spacings (ICLR 2024 · similarity 0.88)
- Temporal Spiking Neural Networks with Synaptic Delay for Graph Reasoning (ICML 2024 · similarity 0.88)
- Advancing Spiking Neural Networks for Sequential Modeling with Central Pattern Generators (NeurIPS 2024 · similarity 0.87)
- TS-SNN: Temporal Shift Module for Spiking Neural Networks (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.