Long Short-Term Memory (LSTM) Spectrum Prediction

LSTMs are architected specifically to solve the vanishing gradient problem that plagues standard RNNs. Through a constant error carousel mechanism, the LSTM cell state allows gradients to flow backward across hundreds or thousands of time steps without decaying. This enables the network to learn dependencies between a current spectrum occupancy state and an event that occurred much earlier, such as a diurnal usage pattern initiated 24 hours prior. The gating structure—comprising input, forget, and output gates—regulates this information flow, learning precisely when to store, discard, or read from the long-term memory cell.

LSTM models can be configured for sequence-to-sequence prediction, outputting a vector of future occupancy states rather than a single point. This is critical for proactive spectrum access, where a cognitive radio needs a forecast for the entire prediction horizon.

• Direct Multi-Output: A dense layer maps the final hidden state to multiple future time steps simultaneously.
• Autoregressive Decoding: The model predicts one step ahead, then feeds its own prediction back as input to generate the next step, iterating to the desired horizon.
• Encoder-Decoder: A separate decoder LSTM is conditioned on the final state of the encoder, generating the forecast sequence one element at a time.

The three gates of an LSTM cell directly model the dynamics of spectrum occupancy:

• Forget Gate: Learns to reset the cell state when a primary user's transmission session ends, discarding now-irrelevant historical context.
• Input Gate: Activates to store new information when a previously idle channel becomes occupied, updating the long-term state with the new signal onset.
• Output Gate: Controls what portion of the cell state is exposed to the next layer, filtering out noise and exposing only the relevant occupancy features for the current time step.

This gating provides a learned mechanism for handling the abrupt state transitions characteristic of bursty wireless traffic.

Real-world spectrum monitoring often produces irregularly sampled time series due to sensor dropouts, varying sweep times, or event-triggered recording. Standard LSTM cells assume uniform time deltas between observations. To address this, Time-Aware LSTM variants incorporate the elapsed time between consecutive measurements as an explicit input feature. The forget gate is modulated by this time delta, allowing the model to decay its memory proportionally to the duration of a gap in observations. This is essential for fusing data from heterogeneous sensing networks with different sampling rates.

A Bidirectional LSTM processes the input sequence in both forward and reverse temporal directions, maintaining two separate hidden states that are concatenated at each time step. For offline spectrum dataset analysis, this allows the model to condition its prediction of an occupancy state on both past and future context within the training window. This is particularly powerful for imputation tasks—filling in missing sensor readings—and for post-hoc anomaly detection where the entire signal trace is available for analysis, providing a more complete temporal context than a unidirectional model.

Standard LSTMs operate on one-dimensional time series. To model correlations across both time and frequency, the Convolutional LSTM replaces matrix multiplications in the gate computations with convolution operations. This allows the model to process a spectrum occupancy matrix as a sequence of 2D images, learning how spectral activity propagates across adjacent frequency channels over time. A ConvLSTM can simultaneously predict that when a wideband signal appears in one channel, adjacent channels will likely become occupied in the next time step, capturing the spectral leakage and bandwidth patterns of real transmissions.