LSTM Spectrum Predictor

The core innovation of the LSTM is the memory cell (C_t), which acts as a conveyor belt of information across time steps. The Constant Error Carousel (CEC) mechanism allows gradients to flow unchanged across many time steps during backpropagation, solving the vanishing gradient problem that plagues standard RNNs. This enables the predictor to learn dependencies spanning hundreds of milliseconds of spectrum activity, such as a primary user's periodic beacon interval.

• Write operation: Controlled by the input gate
• Read operation: Controlled by the output gate
• Reset operation: Controlled by the forget gate

Three distinct, learned gates regulate information flow into and out of the memory cell, each implemented as a sigmoid neural network layer outputting values between 0 and 1.

• Forget Gate (f_t): Decides what historical occupancy data to discard from the cell state. A value near 0 erases a past channel state, while 1 preserves it.
• Input Gate (i_t): Determines which new spectrum observations (x_t) are worth storing in the cell state, often paired with a candidate value (g_t) generated by a tanh layer.
• Output Gate (o_t): Controls what part of the cell state is exposed as the hidden state (h_t) for the current time step's channel state prediction.

For multi-step prediction horizons, the LSTM is often deployed in an encoder-decoder configuration. The encoder ingests a historical sequence of spectrum occupancy vectors (e.g., 100 past time slots) and condenses them into a fixed-length context vector representing the learned temporal dynamics. The decoder then unrolls this context to autoregressively generate a future sequence of channel states (e.g., the next 20 time slots).

• Teacher forcing: During training, the decoder receives the ground-truth previous channel state, accelerating convergence.
• Inference mode: The decoder feeds its own previous prediction as input, enabling autonomous multi-step forecasting for target channel reservation.

Standard LSTMs produce point estimates of future channel occupancy, which are insufficient for risk-aware spectrum handoff decisions. Bayesian LSTMs place a probability distribution over the network weights instead of learning fixed values. By performing multiple stochastic forward passes with Monte Carlo Dropout at inference time, the predictor outputs a predictive distribution with a mean and variance.

• Prediction interval: A 95% confidence interval around the forecasted idle probability
• Risk-averse handoff: The cognitive radio can avoid channels where the predictive variance is high, indicating an unfamiliar or unstable primary user traffic pattern

In offline spectrum survey applications, a Bidirectional LSTM (BiLSTM) processes the time series in both forward and backward directions. This architecture uses two separate hidden layers that scan past-to-future and future-to-past, concatenating their hidden states. This provides the predictor with complete context around each time step, significantly improving interpolation accuracy for filling gaps in historical spectrum measurement data.

• Forward layer: Captures causal dependencies from past observations
• Backward layer: Captures dependencies from future observations relative to the current point
• Use case: Post-mission SIGINT analysis and spectrum usage map generation

To prevent the encoder-decoder bottleneck where a single fixed-length context vector loses information from early time steps, an attention mechanism is added. Instead of relying solely on the final encoder hidden state, the decoder computes a weighted sum of all encoder hidden states at each prediction step. The alignment scores allow the decoder to dynamically focus on the most relevant historical time windows—such as a specific interference burst 50 time steps ago—when predicting the channel state at t+10.

• Bahdanau attention: Additive scoring function, learnable alignment model
• Luong attention: Multiplicative scoring, computationally faster for long sequences

A sequence-to-sequence architecture that maps an input history of spectrum observations to a future sequence of channel states. The encoder compresses the temporal context into a fixed-length vector, while the decoder generates multi-step predictions for target channel reservation. This architecture is fundamental for proactive handoff, allowing a cognitive radio to forecast occupancy several time slots ahead and reserve an idle channel before the current one becomes occupied.

A stochastic framework representing the temporal behavior of licensed spectrum users. Common models include:

• ON/OFF exponential models for voice traffic
• Markovian arrival processes for bursty data
• Markov Modulated Poisson Process (MMPP) for capturing traffic bursts The accuracy of an LSTM predictor depends heavily on how well the training data reflects the underlying PU activity distribution.

A handoff strategy where a secondary user predicts future channel occupancy and switches channels before a primary user arrives. This minimizes service disruption time and reduces forced termination probability. An LSTM Spectrum Predictor enables proactive handoff by forecasting the spectrum availability window for each candidate channel, allowing the cognitive radio to schedule a seamless transition.

A statistical model that infers unobservable channel occupancy states from observable signal measurements. HMMs use Bayesian inference to estimate the hidden state sequence and are a classical baseline for spectrum prediction. LSTMs often outperform HMMs by learning long-range dependencies without requiring explicit specification of a transition probability matrix, making them more adaptable to complex, non-Markovian PU traffic patterns.

An online learning mechanism that detects and adjusts to statistical changes in primary user traffic patterns over time. Without adaptation, an LSTM predictor trained on historical data degrades as PU behavior evolves. Techniques include:

• Sliding window retraining
• Exponential forgetting factors
• Change point detection to trigger model updates This ensures sustained prediction accuracy in dynamic spectrum environments.

A non-parametric Bayesian method providing a predictive distribution over future channel idle times, including a prediction confidence interval. Unlike LSTMs, which output point estimates, Gaussian Processes quantify uncertainty explicitly. This is valuable for risk-aware spectrum handoff, where a secondary user can avoid channels with high prediction variance, even if the mean idle time appears favorable.