Abstract
Spiking Neural Networks (SNNs) offer a promising, biologically inspired approach for processing spatiotemporal data, particularly for time series forecasting. However, conventional neuron models like the Leaky Integrate-and-Fire (LIF) struggle to capture long-term dependencies and effectively process multi-scale temporal dynamics. To overcome these limitations, we introduce the Temporal Segment Leaky Integrate-and-Fire (TS-LIF) model, featuring a novel dual-compartment architecture. The dendritic and somatic compartments specialize in capturing distinct frequency components, providing functional heterogeneity that enhances the neuron's ability to process both low- and high-frequency information. Furthermore, the newly introduced direct somatic current injection reduces information loss during intra-neuronal transmission, while dendritic spike generation improves multi-scale information extraction. We provide a theoretical stability analysis of the TS-LIF model and explain how each compartment contributes to distinct frequency response characteristics. Experimental results show that TS-LIF outperforms traditional SNNs in time series forecasting, demonstrating better accuracy and robustness, even with missing data.
Introduction
Spiking Neural Networks (SNNs) represent an energy-efficient alternative to traditional deep learning by mimicking the discrete electrical pulses used in biological brain communication. However, applying SNNs to time series forecasting has been challenging. The standard Leaky Integrate-and-Fire (LIF) neuron model possesses a "leaky" memory that decays rapidly, making it difficult to capture long-term trends. Additionally, real-world time series data is multi-scale, containing both slow-moving seasonal trends (low frequency) and rapid fluctuations (high frequency). Standard LIF neurons struggle to process these disparate signals simultaneously, often leading to significant information loss.
Methods
The Temporal Segment Leaky Integrate-and-Fire (TS-LIF) model addresses these issues through a dual-compartment architecture that mimics biological neural structures.
As shown in Figure 1, the model separates signal processing into two specialized branches:
- Dendritic Compartment: Acts as a low-pass filter, focusing on slow-changing, long-term trends.
- Somatic Compartment: Acts as a high-pass filter, capturing rapid fluctuations and short-term changes.
- Shortcut Mechanisms: Direct somatic current injection prevents high-frequency details from being washed out during intra-neuronal transmission.
- Dendritic Spiking: The model allows dendrites to generate their own spikes, improving the extraction of multi-scale features.
Beyond the architecture, the authors provide a rigorous theoretical stability analysis and use Z-transforms to prove how each compartment responds to different signal frequencies.
Experimental Evaluation
The model was evaluated on four widely recognized multivariate time series benchmarks: Metr-la and Pems-bay (traffic speed), Solar (power production), and Electricity (consumption). These datasets test the model's ability to handle periodic patterns and sudden, unpredictable shifts.
Key Results
The TS-LIF model was integrated into TCN, GRU, and Transformer backbones, consistently outperforming both traditional SNNs and standard Artificial Neural Networks (ANNs). As detailed in the results table (Figure 5):
- Accuracy: On the Electricity dataset, TS-GRU and TS-former showed a reduction in Root Relative Squared Error (RSE) of up to 43.6% for long-term (96-step) predictions compared to standard spiking models.
- Robustness: In tests with 80% missing data, TS-LIF maintained high accuracy while baseline models saw their performance collapse.
- Efficiency: TS-LIF achieved energy reductions of 60% to 75% compared to non-spiking ANNs.
- Ranking: Across all configurations, TS-LIF variants achieved the highest "Average Rank" among the tested models.
Conclusion
TS-LIF advances the application of SNNs in time-series forecasting by introducing functional heterogeneity and information preservation mechanisms. While the model excels at temporal modeling, future work will focus on capturing cross-variable correlations in multivariate systems and optimizing the framework for deployment on specialized neuromorphic hardware.
Summary updated